Robotic Surveillance and Deployment Strategies

Transcription

Robotic Surveillance and Deployment Strategies
UNIVERSITY OF CALIFORNIA
Santa Barbara
Robotic Surveillance and Deployment Strategies
A Dissertation submitted in partial satisfaction
of the requirements for the degree of
Doctor of Philosophy
in
Mechanical Engineering
by
Rushabh Patel
Committee in Charge:
Professor Francesco Bullo, Chair
Professor Jo˜ao Hespanha
Professor Jeff Moehlis
Professor Bassam Bamieh
Professor Yasamin Mostofi
June 2015
The Dissertation of
Rushabh Patel is approved:
Professor Jo˜ao Hespanha
Professor Jeff Moehlis
Professor Bassam Bamieh
Professor Yasamin Mostofi
Professor Francesco Bullo, Committee Chairperson
April 2015
Robotic Surveillance and Deployment Strategies
c 2015
Copyright by
Rushabh Patel
iii
To my sister Tulsi.
iv
Acknowledgements
First and foremost, I would like to thank my advisor, Francesco Bullo. He
has undoubtably played a critical role in my development as a researcher, and my
accomplishments would not be possible without him. I would especially like to
thank him for his support and understanding through the various ups and downs
that life can bring; it is through that support which my successes were possible.
I know the lessons learned with him, both academic and personal, will have an
everlasting impact on all my future endeavors.
I would also like to give a special thanks to Paolo Frasca, who was an incredible
asset in the early stages of my research. He was very patient with me and acted
as an exceptional sounding board to get my bearings during my first projects. In
addition, his attention to detail helped to make our works that much better.
All of the works that appear in this thesis are the result of various rewarding collaborations with other talented scientist. My thanks go to Joey Durham,
Andrea Carron, Pushkarini Agharkar and Ruggero Carli for all their help. An
additional thanks goes to Pushkarini for her initial discover and discussions of the
Kemeny constant (hitting time), a topic which has proven to be fruitful and a
major focus of this thesis.
During my stay at UCSB, I have had several opportunities that have allowed
me to branch outside my research. My thanks goes to Wendy Ibsen for running
v
the School for Scientific Thought as well as various other outreach programs which
have allowed me to share my love of science with the community. Thank you to
Jeff Peters, who played a critical role in teaching our first high-school robotics
course as well as in the publication of the text which was a direct result of that
course.
I would also like to thank all the present and past members of the Bullo lab.
The many stimulating discussions within the lab, and social gatherings outside of
it, have made my experience at UCSB a truly great one.
Lastly, I would like to thank my friends and family whose supportive words
and understanding have made this achievement possible.
vi
Curriculum Vitæ
Rushabh Patel
Education
2015
Ph.D. Mechanical Engineering
University of California, Santa Barbara,
Santa Barbara, CA, USA
2007
M.S., B.S. Aerospace Engineering
California Polytechnic State University,
San Luis Obispo, CA, USA
Research Experience
2010 – 2015
Graduate Student Researcher, University of California, Santa
Barbara.
Two problems in multi-agent control have been investigated.
Specifically, problems related to (1) coordinated deployment
and partitioning in robotic networks to provide optimal coverage, and (2) unpredictable surveillance and intruder detection
strategies in robotic networks.
2006 – 2007
Master’s Thesis, California Polytechnic State University, San
Luis Obispo.
Investigated the design of adaptive controllers for a four reaction wheel spacecraft simulator.
Teaching Experience
Spring 2014
Teaching Assistant
Course: ME 179P, Introduction to Robotics: Planning and Kinematics
Mechanical Engineering, University of California, Santa Barbara
vii
Spring 2013
Teaching Assistant
Course: ME 155A, Control System Design
Mechanical Engineering, University of California, Santa Barbara
Fall/
Winter 2013
Instructor
Course: Teaching Robotics
School for Scientific Thought, University of California, Santa
Barbara
Winter 2013
Teaching Assistant
Course: ME/ECE 236, Nonlinear Control Systems
Mechanical Engineering, University of California, Santa Barbara
Winter 2011
Teaching Assistant
Course: ME 6, Basic Electrical and Electronic Circuits
Mechanical Engineering, University of California, Santa Barbara
Fall 2010
Teaching Assistant
Course: ME 140A, Numerical Analysis in Engineering
Mechanical Engineering, University of California, Santa Barbara
Mentoring
Summer 2014
Student: Dillon Azzam
Program: Research Mentorship Program
University of California, Santa Barbara
Summer 2014
Student: Nicolae Christoffersen
Program: Research Mentorship Program
University of California, Santa Barbara
Summer 2013
Student: Ariana Del Toro
Program: RISE Summer Internship
University of California, Santa Barbara
Summer 2009
Student: Christopher Walsvick
Program: Northrop Grumman High School Summer Internship
Northrop Grumman Aerospace Systems
viii
Professional Service
Technical Reviewer
Journal
IEEE Transaction on Automatic Control
IEEE Transaction on Control of Network Systems
Automatica
International Journal of Control
Conference
American Control Conference
IEEE Conference on Decision and Control
IEEE Multi-conference on Systems and Control
AIAA Guidance, Navigation and Control Conference
Publications
Texts and Thesis
1. J. R. Peters and R. Patel. Thinking Robotics: Teaching Robots to Make
Decisions. www.teachengineering.org, February 2015. Note: Submitted.
2. R. Patel. Adaptive output feedback control as applied to the rigid body
equations of motion. Master’s thesis, California Polytechnic State University,
San Luis Obispo, December 2007.
Journal Articles
1. R. Patel, A. Carron, and F. Bullo. The Hitting Time of Multiple Random
Walks with Application to Robotic Surveillance. SIAM Journal on Matrix
Analysis and Applications, March 2015. Note: Submitted.
2. R. Patel, P. Agharkar, and F. Bullo. Robotic Surveillance and Markov Chains
with Minimal First Passage Time. IEEE Transactions on Automatic Control,
May 2014. Note: To appear.
3. R. Patel, P. Frasca, and F. Bullo. Centroidal Area-Constrained Partitioning
for Robotic Networks. ASME Journal of Dynamic Systems, Measurement,
and Control, 136(3):031024, 2014.
ix
4. R. Patel, P. Frasca, J. W. Durham, R. Carli, and F. Bullo. Dynamic Partitioning and Coverage Control with Asynchronous One-To-Base-Station Communication. IEEE Transactions on Control of Network Systems, January
2014. Note: To appear.
Conference Articles
1. P. Agharkar, R. Patel, and F. Bullo. Robotic surveillance and Markov chains
with minimal first passage time. In IEEE Conf. on Decision and Control,
Los Angeles, CA, USA, pages 6603-6608, December 2014.
2. R. Patel, P. Frasca, and F. Bullo. Centroidal area-constrained partitioning
for robotic networks. In ASME Dynamic Systems and Control Conference,
Stanford, CA, USA, October 2013.
3. C. Mau-Song, J.C. Donovan, R.C. Patel, and J.K. McCarthy. VIIRS nearfield response, ghosting, and static electrical crosstalk maps methodology. In
Calibration Conference, Honolulu, HI, August 2008.
4. R. Patel and E. Mehiel. Adaptive output feedback control as applied to the
rigid body equations of motion. In AIAA Guidance Navigation and Control
Conference and Exhibit, Logan, UT, August 2008.
x
Abstract
Robotic Surveillance and Deployment Strategies
Rushabh Patel
Autonomous mobile systems are becoming more common place, and have the
opportunity to revolutionize many modern application areas. They include, but
are not limited to, tasks such as search and rescue operations, ad-hoc mobile
wireless networks and warehouse management; each application having its own
complexities and challenging problems that need addressing. In this thesis, we
explore and characterize two application areas in particular.
First, we explore the problem of autonomous stochastic surveillance. In particular, we study random walks on a finite graph that are described by a Markov
chain. We present strategies that minimize the first hitting time of the Markov
chain, and look at both the single agent and multi-agent cases. In the single agent
case, we provide a formulation and convex optimization scheme for the hitting
time on graphs with travel distances. In addition, we provide detailed simulation
results showing the effectiveness of our strategy versus other well-known Markov
chain design strategies. In the multi-agent case, we provide the first characterization of the hitting time for multiple random walkers, which we denote the group
hitting time. We also provide a closed form solution for calculating the hitting
xi
time between specified nodes for both the single and multiple random walker cases.
Our results allow for the multiple random walks to be different and, moreover,
for the random walks to operate on different subgraphs. Finally, we use sequential quadratic programming to find the transition matrices that generate minimal
group hitting time.
Second, we consider the problem of optimal coverage with a group of mobile agents. For a planar environment with an associated density function, this
problem is equivalent to dividing the environment into optimal subregions such
that each agent is responsible for the coverage of its own region. We study this
problem for the discrete time and space case and the continuous time and space
case. For the discrete time and space case, we present algorithms that provide
optimal coverage control in a non-convex environment when each robot has only
asynchronous and sporadic communication with a base station. We introduce the
notion of coverings, a generalization of partitions, to do this. For the continuous
time and space case, we present a continuous-time distributed policy which allows a team of agents to achieve a convex area-constrained partition in a convex
workspace. This work is related to the classic Lloyd algorithm, and makes use
of generalized Voronoi diagrams. For both cases we provide detailed simulation
results and discuss practical implementation issues.
xii
Professor Francesco Bullo
Dissertation Committee Chair
xiii
Contents
Acknowledgements
v
Curriculum Vitæ
vii
Abstract
xi
List of Figures
xvii
List of Tables
xx
1 Introduction
1.1 Literature review . . . . . . . . . . . . . .
1.1.1 Robotic surveillance and the hitting
1.1.2 Partitioning and coverage control .
1.2 Contributions . . . . . . . . . . . . . . . .
1
4
4
8
10
. . .
time
. . .
. . .
. . . . . . .
of a random
. . . . . . .
. . . . . . .
. . .
walk
. . .
. . .
2 The Weighted Hitting Time of a Random Walk
2.1 Notation and Markov chains . . . . . . . . . . . . . . . . . . . .
2.2 The hitting time of a Markov chain and its minimization . . . .
2.2.1 The hitting time for a weighted graph . . . . . . . . . . .
2.2.2 SDP framework for optimizing the hitting time . . . . .
2.3 The weighted hitting time of Markov chain and its minimization
2.3.1 The hitting time for a doubly-weighted graph . . . . . .
2.3.2 SDP framework for optimizing the weighted hitting time
2.3.3 Minimizing single hop distance . . . . . . . . . . . . . .
2.4 Applications of the hitting time to surveillance . . . . . . . . . .
2.4.1 Optimal strategy for Scenario I . . . . . . . . . . . . . .
2.4.2 Numerical analysis of Scenario II . . . . . . . . . . . . .
xiv
.
.
.
.
.
.
.
.
.
.
.
16
16
18
19
25
27
29
35
37
38
40
41
2.5
2.6
Summary . . . . . . . . . . . . .
Proofs and supplemental material
2.6.1 Proof of Theorem 8 . . . .
2.6.2 Supplemental Material . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
3 The Hitting Time of Multiple Random Walks
3.1 Tensor notation and the Kronecker product . . .
3.2 The hitting time of a Markov chain . . . . . . . .
3.2.1 The hitting time of a Markov chain . . . .
3.2.2 Hitting time for doubly-weighted graphs .
3.3 Group hitting time of multiple Markov chains . .
3.3.1 Random-walkers covering the full graph . .
3.3.2 Random-walkers covering subgraphs . . . .
3.3.3 Computational Complexity . . . . . . . . .
3.4 Numerical optimization of the group hitting time
3.4.1 Random-walkers covering the full graph . .
3.4.2 Random-walkers covering subgraphs . . . .
3.4.3 Implementation Notes . . . . . . . . . . .
3.5 Summary . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
47
48
48
51
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
55
56
59
60
66
67
68
77
82
84
86
89
92
93
4 Partitioning with One-to-Base-Station Communication
4.1 Preliminaries and problem statement . . . . . . . . . . .
4.1.1 Graphs and Distances . . . . . . . . . . . . . . .
4.1.2 Coverings of Graphs . . . . . . . . . . . . . . . .
4.1.3 One-to-Base-Station Robotic Network Model . . .
4.1.4 Problem Statement . . . . . . . . . . . . . . . . .
4.2 Proposed Solution . . . . . . . . . . . . . . . . . . . . . .
4.2.1 Cost Functions . . . . . . . . . . . . . . . . . . .
4.2.2 The One-to-Base Coverage Algorithm . . . . . . .
4.2.3 Convergence Proofs . . . . . . . . . . . . . . . . .
4.3 Implementation and Simulations . . . . . . . . . . . . . .
4.3.1 Handling Dynamic Changes . . . . . . . . . . . .
4.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
94
95
95
96
98
99
100
100
104
111
119
124
126
5 Partitioning with Area-Constraints
5.1 Preliminaries and problem statement . . . . . . . .
5.2 Relevant partial derivatives . . . . . . . . . . . . .
5.3 Area-constrained Voronoi partitions . . . . . . . . .
5.4 Centroidal area-constrained Voronoi partitions . . .
5.5 Simultaneous change of agent positions and weights
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
128
129
134
139
146
150
xv
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
5.6
5.7
Implementation and simulations . . . . . . . . . . . . . . . . . . . 152
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
6 Conclusion and Future Work
157
6.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
6.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
Bibliography
163
xvi
List of Figures
2.1 Example of a doubly-weighted graph G = (V, E, P, W ) with three
nodes: (a) shows the edge set, E, allowed for the graph with three nodes,
(b) shows the probabilities, pi,j to move along each edge, and (c) shows
the time (i.e., distance traveled), ωi,j to move along each edge. . . . . .
2.2 Environment with two obstacle represented by an unweighted graph.
43
2.3 Various airport hub locations (top), and the corresponding weight
map (bottom). Edge weights between two hubs account for travel time
between hubs plus required service time once at hub. Self loops have no
travel time so encompass only service time required at hub. . . . . . .
2.4 Percentage of intruders detected for varying intruder life-times by
a surveillance agent executing a random walk according to the Markov
chain generated by the hitting time algorithm (circle), FMMC algorithm
(square), M-H algorithm (asterisk), and the Markov chain generated by
solving Problem 5 (diamond). Average points and standard deviation
error bars are taken over 200 runs, where the intruder appears 500 times
for each run. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.1 From left to right, example of a 5 node ring, 5 node complete, 9
node lattice and 4 node ring graph with self-loops. . . . . . . . . . . .
3.2 Probability to move along each edge of a ring graph (left) and
complete graph (right) for 3 random walkers. In the case above, the
probability to move along each edge is 1 which indicates a cycle. The
group hitting time for the shown trajectories for both ring and complete
graph are H3 = 1.8 . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
xvii
28
45
46
86
87
3.3 Probability to move along each edge of a lattice graph for 1 random
walker (left), 2 random walkers (middle) and 3 random walkers (right).
In each graph, the opacity of a line indicates the probability to move
along an edge. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.4 Probability to move along each edge of a 5 node ring graph with
two agents (left), and 9 node lattice graph with two agents (right). In
each graph, the opacity of a line indicates the probability to move along
an edge. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.5 Probability to move along each edge of a 4 node ring graph with
two agents (left), and 9 node lattice graph with three agents (right). In
each graph, the opacity of a line indicates the probability to move along
an edge. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.1 The left image shows a grid environment whose corresponding
graph representation is shown in the right image. Each cell in the grid
represents a node in the graph and if two cells are adjacent, then there
is an unit-weight edge between those nodes. The black nodes in the
graph denote the set of generalized centroids for the corresponding grid
environment. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2 The figure shows two environments with two agents. Each cell
denotes a node in a graph and if two cells are adjacent, then there is a
unit-weight edge between those nodes. The left image shows a Voronoi
partition generated by the two agents. Note that the blue agent is not
at its region’s centroid. The right image is instead a centroidal Voronoi
partition. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.3 The figure shows three environments with two agents. Each cell
denotes a node in a graph, and if two cells are adjacent then there is an
unit-weight edge between those nodes. . . . . . . . . . . . . . . . . . .
4.4 Snapshots from a simulation of three robots partitioning an environment with black obstacles using the Hmin,inf One-to-base station
algorithm. The free space of the environment is modeled using the indicated occupancy grid where each cell is a vertex in the resulting graph.
The robots’ optimal coverage position is marked by an X and the boundary of each robot’s territory drawn in its color. Some cells are on the
boundary of multiple territories: for these we draw superimposed robot
colors. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.5 Snapshots from a simulation of three robots partitioning an environment with black obstacles using the Hmin,inf One-to-base station
algorithm. Note that the initial condition is the same as in Figure 4.4,
but the evolution is different. . . . . . . . . . . . . . . . . . . . . . . .
xviii
89
90
91
101
104
108
122
123
5.1 The image on the left is the Standard Voronoi Partition generated
by nodes A through E. The image on the right shows the dual graph
for this partition. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2 Simulation of 9 agents partitioning a square environment with uniform density using the iterative gradient algorithm. . . . . . . . . . . .
5.3 Simulation of 9 agents partitioning a square environment with uniform density using the simultaneous gradient algorithm. . . . . . . . .
5.4 Area (left) and position (right) trajectories for 9 agents partitioning a square environment with uniform density using the simultaneous
gradient algorithm. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
xix
132
154
155
156
List of Tables
2.1
runs
2.2
runs
Statistics on the percentage of intruders caught in 200 simulation
for the environment in Fig. 2.2. . . . . . . . . . . . . . . . . . . .
Statistics on the percentage of intruders caught in 200 simulation
for the environment in Fig. 2.3. . . . . . . . . . . . . . . . . . . .
3.1 Group hitting time values for random walks shown in Figure 3.3.
The last column indicates the group hitting time for each case whereas
the middle three columns indicate each random-walkers individual hitting time. Surprisingly, the ring and complete graph exhibit equivalent
hitting time results. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 Group hitting time values for random walks shown in Figure 3.3.
The last column indicates the group hitting time for each case whereas
the middle three columns indicate each random-walkers individual hitting time. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.1 Multi-center function cost-to-cover statistics for each algorithm
from 100 simulation runs. . . . . . . . . . . . . . . . . . . . . . . . . .
xx
43
46
87
88
124
Chapter 1
Introduction
You must construct additional pylons.
- StarCraft (1998)
The applications of multi-agent systems to accomplish complex tasks in a
complex environment are vast. Environmental monitoring [78], search and rescue
operations [58] and ad-hoc mobile wireless networks [38] are just a few areas where
multi-agent systems can have significant impacts. A common problem which
arises in many robotic network applications is that of dividing workload amongst
agents, so tasks can be accomplished quickly and efficiently. The way workload is
distributed amongst a multi-agent platform poses unique challenges for different
application areas. For example, in ocean surveillance the cost of travel is large,
so instead of each agent traveling around the entire region of interest, it is more
effective for each agent to survey small subsets of the larger region. Placement of
agents within such regions can also be of importance. In the case of surveillance or
warehouse management, it is desirable to be at the center of your region so tasks
1
Chapter 1. Introduction
are more easily serviced. Simultaneous placement of agents and specification of
those agents’ regions is also of general interest, however, this can be complicated
as the two problems are often coupled. In this thesis we study two closely related
robotic applications; robotic surveillance and robotic partitioning and coverage
control.
Robotic surveillance The design of surveillance algorithms for quickest detection of intruders and anomalies appear frequently. Specific examples include the
monitoring of oil spills [20], the detection of forest fires [48], the tracking of border
changes [81], and the periodic patrolling of an environment [30, 68]. Other applications in single and multi-agent systems include minimizing emergency vehicle
response times [7] as well as servicing tasks in robotic warehouse management [86].
In areas of research outside of robotics, applications include, but are not limited
to, determining how quickly epidemics spread [85], how information propagates in
a social network [5] and how quickly information packets get transferred in a wireless node network [75]. In this paper we propose stochastic surveillance strategies
based on Markov chains. More specifically, we look at the analysis, generalization
and minimization of the hitting time of a random walk governed by a Markov
chain for both single and multiple random walkers. This problem is not only of
interest in the context of robotic surveillance, but is also of general mathematical
2
Chapter 1. Introduction
interest in the study of Markov chains and random walks; similar to the fastest
mixing Markov chain, the first hitting time is a metric by which to gauge the
performance of a random walk.
Coverage control In applications such as environmental monitoring or warehouse logistics a team of robots is asked to perform tasks over a large space.
The distributed environment partitioning problem consists of designing control
and communication laws for individual robots such that the team divides a space
into regions in order to optimize the quality of service provided. Coverage control
additionally optimizes the positioning of robots inside of a region. Coverage control and territory partitioning have applications in many fields. In cyber-physical
systems, applications include automated environmental monitoring [33], fetching
and delivery [86], and other vehicle routing scenarios [35]. In this thesis we consider two variations of the partitioning problem. In the first problem we provide a
method for partitioning a discretized environment using an asynchronous one-tobase station communication law. In the second problem we look a the partitioning
of continuous environment in which each agent only has communication with its
neighbor. In both cases we consider different multi-center cost functions and determine optimal partitions in relation to those cost functions.
3
Chapter 1. Introduction
1.1
1.1.1
Literature review
Robotic surveillance and the hitting time of a random walk
In this paper we study strategies to surveil an environment, to provide a desired
coverage frequency, and to detect an intruder in minimum time. The surveillance
problem has appeared in the literature in various manifestations. The authors
of [80] look at minimizing time of detection of noisy anomalies via persistent
surveillance strategies, and in [55] wireless sensor networks are utilized for intruder detection in previously unknown environments. In [4], the authors explore
strategies for surveillance using a multi-agent ground vehicle system which must
maintain connectivity between agents. A non-cooperative game framework is utilized in [19] to determine an optimal strategy for intruder detection, and in [69]
a similar framework is used to analyze intruder detection for ad-hoc mobile networks. In our setup we model the environment as a graph and design random walks
on this graph imposing restrictions on the stationary distribution of the walk. In
other works with a similar setup, Markov chain Monte Carlo methods [36, 79]
are used to design surveillance strategies. In [36] convexity results for symmetric
matrices are utilized to further optimize those strategies. Deterministic policies
have been used to minimize the visit frequencies in a graph [29, 80], however, a
4
Chapter 1. Introduction
main result of [79] shows that deterministic policies are ill-suited when designing
strategies with arbitrary constraints on those visit frequencies. We take an alternate approach and design policies using Markov chains with minimal hitting
time.
The hitting time of a random walk governed by a Markov chain, is the expected
time taken by a random walker to travel between any two nodes in a network. For
a single finite discrete-time Markov chain, this quantity is also well-known as the
Kemeny constant of the Markov chain, the mean first passage time of a Markov
chain or, for the case of reversible Markov chains, the eigentime of a Markov
chain. We refer to this quantity as the first-hitting time or simply hitting time
of a Markov chain, due both to the descriptive nature of this coinage as well as
its prevalence in the literature. The hitting time of a Markov chain for a finite
irreducible Markov chain first appeared in [47], however, it was rediscovered for
finite reversible Markov chains in [11]. Since its original discovery, the hitting
time has been further developed by several groups [45, 49, 70]. The authors
of [45, 53] give bounds on the hitting time for various graph topologies and in [18]
an alternate formulation is explored. Recently, the authors of [70] extended the
notion of the hitting time to networks with travel distances and provide a scheme
for its optimization, and the authors of [2] provided the first formulation of the
hitting time for continuous time reversible Markov chains.
5
Chapter 1. Introduction
The hitting time is closely related to several other well-studied Markov chain
properties. We focus on four quantities that are of particular interest, however,
several others exist. The first and most closely related quantity is the pairwise
hitting time between two nodes, which is the expected time to travel between a
specified pair of nodes. Clearly, the hitting time is simply the expectation of all
possible pairwise hitting times of a Markov chain. Using the relation between
reversible Markov chains and electrical networks [25] the authors of [82] give analytic expressions for hitting times in terms of effective resistance. Using the
electrical framework, closed form expressions for pairwise hitting times have also
been given for special cases of certain Markov chains [67, 66], however, to the
best of our knowledge no general closed form expression exists. Second, the cover
time of a graph is the expected time it takes to reach every node in the graph
at least once. This quantity is sometimes interpreted as a function of a single
node, or more generally, in the context of an arbitrary set of nodes. There are
several works relating the cover time to the hitting time of a transition matrix [61];
many of these works bound the cover time in terms of pairwise hitting time (most
often the worst case pairwise hitting time). Third, the mixing rate of an irreducible Markov chain is the rate at which an arbitrary distribution converges to
the chain’s stationary distribution. The influential text [54] provides a detailed
review of the mixing rate and of other notions of mixing. Recently, [49] refers to
6
Chapter 1. Introduction
the hitting time as the “expected time to mixing” and relates it to the mixing
rate. Finally, the Kirchhoff index [50], also known as the effective graph resistance
[28], is a related metric quantifying the distance between pairs of vertices in an
electric network. The relationship between electrical networks and random walks
on graphs is explained elaborately in [25]. For an arbitrary graph, the Kirchoff
index and the hitting time can be calculated from the eigenvalues of the conductance matrix and the transition matrix, respectively. The relationship between
these two quantities for regular graphs is established in [65].
In this work, we also look at the hitting time of multiple random walkers.
More specifically, we analyze the expected first time to reach any single node in
a network given that there are an arbitrary number of random walkers in that
network. Recently, the authors of [27] look at bounds on hitting times and cover
times for multiple random walkers. In [3] alternate bounds on cover time are
formulated for reversible Markov chains and tight bounds were explored in [31, 32].
The authors of [22, 21] find solutions for cover times in the limit as the number
of nodes in the graph becomes infinite. However, as far as we can discern, a key
assumption made by all work presented thus far in the literature is that results
are based on k copies of a simple random walk over a single graph. We make
no such assumptions in our work; every random walker can move according to
a different and arbitrary random walk and need not share the same underlying
7
Chapter 1. Introduction
graph topology. Also, as opposed to prior work, our results are not bounds but
exact analytic expressions.
To achieve our results, we utilize the notion of Kronecker graphs. Preliminary
results for undirected Kronecker graphs were introduced in [84], showing conditions under which the Kronecker product of two graphs generates a connected
graph. In [51, 52] the authors consider and analyze special Kronecker graphs that
are created by products of the same n × n edge matrix in order to model large
networks and also introduce the concept of “stochastic” Kronecker graphs to generate large networks. This method of generating networks is further refined for
the special case of n = 2 in [59, 76]. In our work we also utilize the notion of
Kronecker products between stochastic matrices, but this should not be confused
with the notion of “stochastic” Kronecker graphs previously mentioned. In this
work we are not attempting to generate network models, but instead are utilizing
novel aspects of Kronecker products and stochastic matrices that have, to the best
of our knowledge, not been deeply explored; the ideas presented here would most
closely be linked to that of [84] of the previously mentioned works.
1.1.2
Partitioning and coverage control
A broad discussion of partitioning and coverage control is presented in [13]
which builds on the classic work of Lloyd [56] on algorithms for optimal quan-
8
Chapter 1. Introduction
tizer design through “centering and partitioning.” The Lloyd-type approach was
first adapted for distributed coverage control in [24] and has since seen many
variations, including non-convex environments [73, 6] and self-triggered coverage
algorithms [62].
Many existing coverage control algorithms assume that robots can communicate peer-to-peer [24], but in some environments this is impractical. For example,
underwater acoustic communication between ocean gliders is very low bandwidth
and hilly or urban terrain can block radio communication. Instead, we present a
coverage control algorithm for a team of robots which collectively maintain complete coverage of the environment and individually have only occasional contact
with a central base station. This one-to-base-station communication model can
represent ocean gliders surfacing to communicate with a tower [72], UAV data
mules that periodically visit ground robots [77], or cost-mindful use of satellite or
cellular communication. Our algorithm optimizes the response time of the team
to service requests in a non-convex environment represented by a graph, with
optimality defined by relevant “multi-center” cost functions for overlapping territories. Early work in coverage control of discrete non-convex domains (represented
by graphs) is presented in [26]. Discrete coverage problems are closely related to
the literature on data clustering and k-means [46], as well as the facility location
or k-center problem [83].
9
Chapter 1. Introduction
This work also utilizes the notion of generalized Voronoi partitions. Results
on specific manifestations of generalized Voronoi partitions and partitioning can
be found in [64]. Results on the existence of power diagrams (a special generalized Voronoi partition manifestation) with area-constraints, along with a method
to determine them are presented in [71]. More detailed results on existence of
generalized Voronoi partitions for arbitrary area constraints are presented in [23].
Linear programming is used to handle generalized Voronoi partitions in [17] for
fixed agents, showing that generalized Voronoi partitions are optimal for a certain
class of multicenter functions. Similar results are obtained in [23], together with
a discrete-time algorithm to solve the problem of optimal deployment of agents,
while satisfying constraints on the areas.
1.2
Contributions
In this section, we highlight the contributions of the thesis organized by chapter.
Chapter 2 Before stating our contributions, it is worth mentioning that all
work to date on the hitting time and mixing rate of a Markov chain on a graph
has been completed under the assumption of homogeneous travel time along the
edges of the graph. The contributions of this chapter are then six fold. First,
10
Chapter 1. Introduction
we provide a convex optimization framework to minimize the hitting time of a
reversible Markov chain given the underlying graph topology of the random walk
and the desired stationary distribution. Second, using doubly-weighted graphs
we extend the formulation of the hitting time to the network environments with
non-homogeneous travel times, a generalization not yet looked at in the literature.
We denote this extension the weighted hitting time. Third, we derive a closed form
solution for the weighted hitting time and show its relation to the hitting time.
Fourth, we provide a convex optimization framework to minimize the weighted
hitting time of a Markov chain with desired stationary distribution. Fifth, we
provide a semidefinite program (SDP) formulation for the optimization of the
hitting time and the weighted hitting time. Finally, we look at two stochastic
surveillance scenarios; in the first scenario we provide a setup in which minimizing
the weighted hitting time leads to the optimal Markov-chain strategy. In the
second surveillance scenario we establish through numerical simulation that the
Markov chain with the minimum weighted hitting time performs substantially
better compared with other well-known Markov chains (i.e., the fastest mixing
chain and the Metropolis-Hastings Markov chain).
Chapter 3 There are several key contributions of this chapter. First, we provide a novel method for the computation of the hitting time of a Markov chain. In
11
Chapter 1. Introduction
the process, we also provide the first closed form method for calculating pairwise
hitting times between any two nodes for an arbitrary irreducible Markov chain.
To the best of our knowledge, results for determining pairwise hitting times until now require restriction to reversible Markov chains, where knowledge of the
Markov chain stationary distribution is known in advance. Our solution has no
such restrictions. Second, we extend the notion of pairwise hitting time to an
irreducible Markov chain with travel distances between nodes. Third, we define
and provide the first closed form solution for computing the hitting time given
multiple (different) Markov chains on the same graph. We denote this extension
the ‘group’ hitting time. Fourth, our results also allow for the extension and calculation of the pairwise hitting time to the hitting time between any set of nodes
for multiple random walkers; for any combination of specified starting nodes, we
can calculate the first hitting time to a specified desired node. Fifth, we further
extend the notion of group hitting time and hitting times between sets of nodes
from multiple random walks on the same graph to random walks on multiple subgraphs. Finally, we provide a detailed numerical analysis to build intuition on the
transition matrices that generate a minimal group hitting time. Before stating the
paper organization, it is worthwhile to note that we achieve our analytical results
by introducing a method of proof that utilizes the Kronecker product; thus our
12
Chapter 1. Introduction
work not only provides results in Markov chain behavior, but also gives general
insight into Kronecker graphs and stochastic matrices.
Chapter 4 The contributions of this chapter are four-fold. First, we present
the first coverage control algorithm for an asynchronous one-to-base-station communication model. This model is realistic and relevant for a variety of application
domains. We handle the time delay between when robots communicate with the
base station using overlapping regions instead of a partition. The algorithm can be
adapted for various cost-functions and also allows for heterogeneity among agents.
Second, we prove that the algorithm converges to a centroidal Voronoi partition
in finite time for two relevant cost-functions. Our Lyapunov argument is based
on an adaptation of the standard partition-based coverage cost function. Third,
we introduce the notion of Pareto-optimal partitions and provide a cost-function
to achieve such a partition using our algorithm. Finally, we describe how the algorithm can seamlessly handle changes in the environment as well as unscheduled
arrival, departure or change in functionality of robots from the team. This feature
leverages overlapping regions, and also eases integration of coverage control with
task servicing.
Chapter 5 The contributions of this chapter are four-fold. First, we design
a provably correct, spatially-distributed continuous-time algorithm to compute
13
Chapter 1. Introduction
area-constrained generalized Voronoi partitions of a convex environment. Our
approach improves upon the work done in [71] for power diagrams both in generality and in numerical stability. Second, we build on work in [23], by introducing a continuous-time spatially-distributed algorithm to compute centroidal areaconstrained generalized Voronoi partitions of a convex environment. Although an
approximate method is proposed in [71] to achieve this for power diagrams, here
we introduce a simpler generalized method which bridges the gap between [71]
and [23] in that it is the continuous-time analogue to [23]. More precisely, the
continuous-time algorithm presented in this paper and the discrete-time algorithm
in [23] both converge to the set of centroidal area-constrained Voronoi partitions,
which shall be formally defined below. Our proposed continuous-time algorithms
are not only of academic interest in their own right, but also fill application gaps
where a discrete-time algorithm may fall short. One such application is in mobile robotic networks, when there is a need to have continuous adaptive coverage
and the understanding of the region’s underlying density distribution changes frequently. For such cases, in the discrete time setting it is possible that, before
the agents arrive at their optimal configuration, the optimal configuration has
changed due to a new understanding of the density distribution. Moreover, the
path that is taken may cause coverage cost to increase temporarily, since only
the configuration is optimal not the path taken. In the continuous-time setting,
14
Chapter 1. Introduction
the agents are always moving in a path that improves their coverage cost, and
a change in the underlying distribution simply means a correction in their path.
Third, we introduce a practical method for implementation of our algorithms,
and show their performance in simulation. Finally, as theoretical contributions,
we prove that the set of mappings that generate area-constrained partitions are
unique up to translation and we compute several novel partial derivatives of relevant operators. Specifically, we generalize a classic result about multi-center
optimization: the partial derivative of the multicenter function evaluated at areaconstrained Voronoi partitions has the same direction as the vector connecting
the robot locations to the centroids of their regions.
15
Chapter 2
The Weighted Hitting Time of a
Random Walk
This chapter is organized as follows. In Section 2.1 we summarize the notation
which we use in this chapter and in Chapter 3, and briefly review properties of
Markov chains. In Section 2.2 we give relevant background for the hitting time and
present our results for its minimization. In Section 2.3 we introduce and provide
detailed characterization of the weighted hitting time as well as its minimization.
In Section 2.4 we provide practical surveillance applications of the weighted hitting
time. In the final section we summarize our findings.
2.1
Notation and Markov chains
We use the notation A = [ai,j ] to denote a matrix A with the element ai,j in its
i-th row and j-th column and, unless otherwise indicated, use bold-faced letters to
16
Chapter 2. The Weighted Hitting Time of a Random Walk
denote vectors. Letting δi,j denote the Kronecker delta, Ad = [δi,j ai,j ] represents
the diagonal matrix whose diagonal elements are the diagonal elements of the
matrix A. The column vector of all ones and length n is denoted by 1n ∈ Rn×1
and I represents the identity matrix of appropriate dimension. We use diag[b] to
denote the diagonal matrix generated by vector b and Tr[A] to denote the trace
of matrix A.
A Markov chain is a sequence of random variables taking value in the finite
set {1, . . . , n} with the Markov property, namely that, the future state depends
only on the present state; see [43, 47] for more details. Let Xk ∈ {1, . . . , n} denote
the location of a random walker at time k ∈ {0, 1, 2, . . . }. We are now ready to
summarize some terminology and results for Markov chains. (1) A Markov chain
is time-homogeneous if P[Xn+1 = j|Xn = i] = P[Xn = j|Xn−1 = i] = pi,j , where
P ∈ Rn×n is the transition matrix of the Markov chain. (2) The vector π ∈ Rn×1
is a stationary distribution of P if
Pn
i=1
π i = 1, 0 ≤ π i ≤ 1 for all i ∈ {1, . . . , n}
and π T P = π T . (3) A time-homogeneous Markov chain is said to be reversible if
π i pi,j = π j pj,i , for all i, j ∈ {1, . . . , n}. For reversible Markov chains, π is always a
steady state distribution. (4) A Markov chain is irreducible if there exists a t such
that for all i, j ∈ {1, . . . , n} , (P t )i,j > 0. (5) If the Markov chain is irreducible,
then there is a unique stationary distribution π, and the corresponding eigenvalues
of the transition matrix, λi for i ∈ {1, . . . , n}, are such that λ1 = 1, |λi | ≤ 1 and
17
Chapter 2. The Weighted Hitting Time of a Random Walk
λi 6= 1 for i ∈ {2, . . . , n}. For further details on irreducible matrices and about
results (4) and (5) see [60, Chapter 8]. In this thesis we consider finite irreducible
time-homogeneous Markov chains.
2.2
The hitting time of a Markov chain and its
minimization
Consider a undirected weighted graph G = (V, E, P ) with node set V :=
{1, . . . , n}, edge set E ⊂ V × V, and weight matrix P = [pi,j ] with the property that pi,j ≥ 0 if (i, j) ∈ E and pi,j = 0 otherwise. We interpret the weight
of edge (i, j) as the probability of moving along that edge. Therefore, element
pi,j in the matrix represents the probability with which the random walk visits
node j from node i. Throughout this document we assume that the underlying
undirected graph (V, E) associated to the transition probabilities P is connected.
In this section we look into a discrete-time random walk defined by a Markov
chain on a graph G. At each time step (hop) of the random walk we move to a new
node or stay at the current node according to the transition probabilities described
by a transition matrix P as discussed above. We do this with three objectives
in mind. The first objective is to analyze the random walk and characterize the
average visit time between nodes in the graph. The second objective is to minimize
18
Chapter 2. The Weighted Hitting Time of a Random Walk
the average visit time between any two nodes and the final is to achieve a long
term (infinite horizon) visit frequency π i at node i. Here, the frequency π i is the
ratio of visits to node i divided by the total number of visits to all nodes in the
graph. Throughout the paper, we describe the random walk using realizations of
a Markov chain with transition matrix P = [pi,j ].
2.2.1
The hitting time for a weighted graph
Let Xk ∈ {1, . . . , n} denote the location of the random walker at time k ∈
{0, 1, 2, . . . }. For any two nodes i, j ∈ {1, . . . , n}, the first passage time from i to
j, denoted by Ti,j , is the first time that the random walker starting at node i at
time 0 reaches node j, that is,
Ti,j = min{k ≥ 1 | Xk = j given that X0 = i}.
It is convenient to introduce the shorthand mi,j = E[Ti,j ], and to define the mean
first passage time matrix M to have entries mi,j , for i, j ∈ {1, . . . , n}. The mean
first passage time from start node i, denoted by hi , is the expected first passage
time from node i to a random node selected according to the stationary distribution π, i.e.,
hi =
n
X
mi,j π j .
j=1
19
Chapter 2. The Weighted Hitting Time of a Random Walk
It is well known [45] that the mean first passage time from a start node is independent of the start node, that is, hi = hj for all i, j ∈ {1, . . . , n}. Accordingly,
we let H = hi , for all i ∈ {1, . . . , n}, denote the hitting time, also known as the
Kemeny constant, of the Markov chain.
Next, we provide formulas for these quantities. By definition, the first passage
time from i to j satisfies the recursive formula:




1,
with probability pi,j ,
Ti,j =



Tk,j + 1, with probability pi,k , k 6= j.
Taking the expectation, we compute
mi,j = pi,j +
n
X
n
X
pi,k (mk,j + 1) = 1 +
pi,k mk,j ,
k=1,k6=j
k=1,k6=j
or in matrix notation,
(I − P )M = 1n 1Tn − P Md ,
(2.1)
where P is the transition matrix of the Markov chain. If the Markov chain is irreducible with stationary distribution π, then one can show Md = diag[{1/π 1 , . . . , 1/π n }],
and
T
π Mπ =
n
X
i=1
πi
n
X
π j mi,j =
j=1
n
X
π i hi = H.
i=1
Clearly, the hitting time can be written as the function P 7→ H(P ), however, to
ease notation we simply write H and use H(P ) only when we wish to emphasize
its dependence on P .
20
Chapter 2. The Weighted Hitting Time of a Random Walk
The hitting time H = π T M π can be computed from equation (2.1) or can be
expressed as a function of the eigenvalues of the transition matrix P as is stated
in the following theorem [45].
Theorem 1. (Hitting time of an irreducible Markov chain): Consider a Markov
chain with an irreducible transition matrix P with eigenvalues λ1 = 1 and λi ,
i ∈ {2, . . . , n}. The hitting time of the Markov chain is given by
H =1+
n
X
i=2
1
.
1 − λi
Using Theorem 1, we derive the following equivalent expression for reversible
Markov chains in terms of the trace of a symmetric positive definite matrix. Before
stating our result, we first introduce some notation. Given a stationary distribution vector π ∈ Rn×1 for a Markov chain with transition matrix P ∈ Rn×n ,
we define the matrix Π ∈ Rn×n as Π = diag[π] and the vector q ∈ Rn×1 as
√
√
q T = ( π 1 , . . . , π n ). We are now ready to state our first result.
Theorem 2. (Hitting time of a reversible irreducible Markov chain): The hitting time of a reversible irreducible Markov chain with transition matrix P and
stationary distribution π is given by
H = Tr (I − Π1/2 P Π−1/2 + qq T )−1 .
21
(2.2)
Chapter 2. The Weighted Hitting Time of a Random Walk
Proof. We start by noting that P is an irreducible row-stochastic matrix therefore
the eigenvalues of P are {λ1 = 1, λ2 , . . . , λn }, where |λi | ≤ 1 and λi 6= 1 for
i ∈ {2, . . . , n}. It follows that the eigenvalues of (I −P ) are {0, 1 − λ2 , . . . , 1 − λn }.
Since P is irreducible and reversible, there exists a stationary distribution π ∈ Rn>0
implying Π is invertible and that Π1/2 (I −P )Π−1/2 = I −Π1/2 P Π−1/2 is symmetric.
It can easily be verified that I − P and I − Π1/2 P Π−1/2 have the same eigenvalues
and that q is the eigenvector associated with the eigenvalue at 0. Next, notice the
matrix (I−Π1/2 P Π−1/2 +qq T ) is symmetric and that (I−Π1/2 P Π−1/2 +qq T )q = q.
Therefore, (I−Π1/2 P Π−1/2 +qq T ) has an eigenvalue at 1 associated with the vector
q. Since (I − Π1/2 P Π−1/2 + qq T ) is symmetric, the eigenvectors corresponding
to different eigenvalues are orthogonal; implying for eigenvector v 6= q that (I −
Π1/2 P Π−1/2 + qq T )v = (I − Π1/2 P Π−1/2 )v since the eigenvalue at 1 is simple.
Therefore, the eigenvalues of (I − Π1/2 P Π−1/2 + qq T ) are {1, 1 − λ2 , . . . , 1 − λn }.
Thus, H = Tr (I − Π1/2 P Π−1/2 + qq T )−1 = 1 + 1/(1 − λ2 ) + . . . + 1/(1 − λn ) =
H.
Given the above result, we are now ready to state our first problem of interest.
Problem 1. (Optimizing the hitting time of a reversible Markov chain): Given the
stationary distribution π and graph G with vertex set V and edge set E, determine
22
Chapter 2. The Weighted Hitting Time of a Random Walk
the transition probabilities P = [pi,j ] solving:
minimize Tr (I − Π1/2 P Π−1/2 + qq T )−1
subject to
n
X
pi,j = 1, for each i ∈ {1, . . . , n}
j=1
π i pi,j = π j pj,i , for each (i, j) ∈ E
(2.3)
0 ≤ pi,j ≤ 1, for each (i, j) ∈ E
pi,j = 0, for each (i, j) ∈
/ E.
Remark 3. All feasible solutions P to Problem 1 are inherently irreducible transition matrices: when P is not irreducible, the matrix (I −Π1/2 P Π−1/2 +qq T ) is not
invertible. Moreover, a feasible point always exists since the Metropolis-Hastings
algorithm applied to any irreducible transition matrix associated with G, generates a reversible transition matrix which is irreducible and satisfies the stationary
distribution constraint [39].
The following theorem establishes the convexity of the hitting time for transition matrices with fixed stationary distribution.
Theorem 4 (Convexity of Problem 1). Let Pπ denote the set of matrices associated to irreducible reversible Markov chains with stationary distribution π. Then,
Pπ is a convex set and P 7→ H(P ) is a convex function over Pπ .
Proof. Let S denote the set of symmetric positive definite matrices, for any stationary distribution π ∈ Rn>0 , denote the set SP,π := {I −Π1/2 P Π−1/2 +qq T | P ∈
23
Chapter 2. The Weighted Hitting Time of a Random Walk
Pπ }. We begin by showing that Pπ is a convex set. Let P1 , P2 ∈ Pπ , then Pπ is
convex if for an arbitrary θ ∈ [0, 1] that
θP1 + (1 − θ)P2 = P3 ∈ Pπ .
(2.4)
Pre and post multiplying (2.4) by Π1/2 and Π−1/2 , respectively, we have that
θΠ1/2 P1 Π−1/2 + (1 − θ)Π1/2 P2 Π−1/2 = Π1/2 P3 Π−1/2 . Then Π1/2 P3 Π−1/2 is symmetric since Π1/2 P1 Π−1/2 and Π1/2 P2 Π−1/2 are symmetric. Pre multiplying (2.4) by
π T we easily verify that the stationary distribution P3 is π T and similarly, post
multiplying by 1n verifies that P3 is row stochastic. Finally taking the left hand
side of (2.4) to the n-th power gives (θP1 + (1 − θ)P2 )n = θn P1n + (1 − θ)n P2n + ζ,
where ζ denotes the sum of all other terms in the expansion and has the property
ζi,j ≥ 0 for all i, j since P1 and P2 are non-negative element-wise matrices. Moreover from irreducibility, there exists a sufficiently large N such that for n > N ,
(P1n )i,j > 0 and (P2n )i,j > 0 for all i, j, which implies (P3n )i,j > 0, therefore P3 ∈ Pπ
and Pπ is convex.
Next we show that SP,π ⊂ S. From the proof of Theorem 2 we have for
P ∈ Pπ that I − Π1/2 P Π−1/2 + qq T has eigenvalues {1, 1 − λ2 , . . . , 1 − λn }, where
λi for i ∈ {1, . . . , n} are the eigenvalues of P , where λi ≤ |1| for all i and λi 6= 1
for i ∈ {2, . . . , n}. Therefore, all eigenvalues of I − Π1/2 P Π−1/2 + qq T are strictly
greater than zero. Finally, since P is reversible Π1/2 P Π−1/2 is symmetric implying
(I − Π1/2 P Π−1/2 + qq T )T = I − Π1/2 P Π−1/2 + qq T and so SP,π ⊂ S.
24
Chapter 2. The Weighted Hitting Time of a Random Walk
Finally, define the mapping g : Pπ 7→ SP,π by g(X) = I − Π1/2 XΠ−1/2 + qq T .
This is an affine mapping from the convex set Pπ to a subset of S. From [34] we
know that Tr[X −1 ] is convex for X ∈ S, therefore the composition with the affine
mapping g : Pπ 7→ SP,π ⊂ S, Tr[g(X)−1 ] is also convex [10, Chapter 3.2.2].
Problem 1 includes constraints on the stationary distribution of the transition
matrix, a notion which has not been looked at in the literature before. The author
of [49] provides bounds to determine the set of transition matrices such that H is
minimized and [45] gives special matrices for which the optimal hitting time can
be found, but these are all approached for the general setting with no constraint
on the actual stationary distribution. In real-world settings, constraints on the
stationary distribution are important and have many practical interpretations.
For example, it is often desirable to visit certain regions more frequently than
other based on each region’s relative importance.
2.2.2
SDP framework for optimizing the hitting time
Here we show how Problem 1 can be equivalently rewritten as an SDP by
introducing a symmetric slack matrix.
Problem 2. (Optimizing the hitting time of a reversible Markov chain (SDP)):
Given the stationary distribution π and graph G with vertex set V and edge set E,
25
Chapter 2. The Weighted Hitting Time of a Random Walk
determine X = [xi,j ] and the transition probabilities P = [pi,j ] solving:
minimize Tr[X]
subject to


 I − Π1/2 P Π−1/2 + qq T I 

0


I
X
n
X
pi,j = 1, for each i ∈ {1, . . . , n}
j=1
π i pi,j = π j pj,i , for each (i, j) ∈ E
0 ≤ pi,j ≤ 1, for each (i, j) ∈ E
pi,j = 0, for each (i, j) ∈
/ E.
The first inequality constraint in Problem 2 represents a linear matrix inequality (LMI) and denotes that the matrix is positive semidefinite. Since the matrix
in the LMI has off-diagonal entries equal to the identity matrix, it holds true if
and only if X is positive definite and (I − Π1/2 P Π−1/2 + qq T ) − X −1 is positive
semidefinite [1, Theorem 1]. This implies (I − Π1/2 P Π−1/2 + qq T ) is positive
definite and that X (I − Π1/2 P Π−1/2 + qq T )−1 . Therefore, the SDP given by
Problem 2 minimizes the hitting time.
26
Chapter 2. The Weighted Hitting Time of a Random Walk
2.3
The weighted hitting time of Markov chain
and its minimization
In most practical applications, distance/time traveled and service costs/times
are important factors in the model of the system. We incorporate these concepts
by allowing for an additional set of weighted edges in our graph in addition to the
edge weights which describe the transition probabilities. Such a system can be
represented by the doubly-weighted graph G = (V, E, P, W ), where W = [ωi,j ] is
a weight matrix with the properties that: if (i, i) ∈ E, then ωi,i ≥ 0; if (i, j) ∈ E,
i 6= j, then ωi,j > 0 ; and if (i, j) ∈
/ E, then ωi,j = 0. The weighted adjacency
matrix P = [pi,j ] has the same interpretation as before as an irreducible rowstochastic transition matrix P which governs the random walk on the graph. An
example of a doubly-weighted graph is shown in Figure 2.1. In the following, we
will interpret ωi,j , i 6= j as the time to travel between two nodes, i and j, in
the graph and ωi,i as the service time at node i. We discuss another motivating
example and interpretation for ωi,j in a later section.
Recall that Xk = i denotes that the random walker is at node i at time k.
If a sample trajectory of the random walk is X0 = i, X1 = j, X2 = k, then the
time instant at which a random walker arrives in state X2 is ωi,j + ωjk . Thus the
27
Chapter 2. The Weighted Hitting Time of a Random Walk
p11
!11
1
1
1
p31
p13
3
p23
3
2
!31
p22
!13
!23
2
3
p32
p33
!22
2
!32
!33
Figure 2.1: Example of a doubly-weighted graph G = (V, E, P, W ) with three
nodes: (a) shows the edge set, E, allowed for the graph with three nodes, (b)
shows the probabilities, pi,j to move along each edge, and (c) shows the time (i.e.,
distance traveled), ωi,j to move along each edge.
time interval between two consecutive steps of this random walk depends on the
weighted adjacency matrix, W , of the graph and is not constant.
In the following analysis, we look at several characterization and optimization
objectives: The first involves extending the notion of the hitting time to doublyweighted graphs and providing a detailed characterization of this extension. The
second involves the minimization of the hitting time of a doubly-weighted graph
and the third involves characterization and minimization of the mean time to
execute a single hop. The first and second objectives are motivated by the need
to minimize visit times to nodes in the graph, and the third is motivated by the
desire to minimize resource consumption when moving between nodes. We seek to
design transition matrices P with stationary distribution π which optimize each
problem. We start with the first objective.
28
Chapter 2. The Weighted Hitting Time of a Random Walk
2.3.1
The hitting time for a doubly-weighted graph
The mean first passage time for the Markov chain on a weighted graph G =
(V, E, P ) by definition, is simply its hitting time. Recall that the mean first
passage time for node i, defined by hi , is determined by taking the expectation
over the first passage times mi,j , from node i to all other nodes j. We present
an analogous notion of the first passage time between two nodes on a doublyweighted graph. We start with defining the first passage time random variable for
a random walk on a doubly-weighted graph and provide a recursive formulation
for its expectation.
As in Section 2.2.1, for any two nodes i, j ∈ {1, . . . , n}, the first passage time
from i to j is the first time that the random walker starting at node i at time 0
reaches node j, that is,
Ti,j = min
k−1
nX
o
wXn ,Xn+1 , for k ≥ 1 | Xk = j given that X0 = i .
n=0
Lemma 5. (First passage time for a doubly-weighted graph): Let mi,j (W ) =
E[Ti,j ] denote the mean first passage time to go from i to j for a graph with weight
matrix W and transition matrix P . Then
mi,j (W ) = pi,j (ωi,j ) +
X
pi,k (mk,j (W ) + ωi,k ),
(2.5)
k6=j
or, in matrix notation,
(I − P )M = (P ◦ W )1n 1Tn − P Md ,
29
(2.6)
Chapter 2. The Weighted Hitting Time of a Random Walk
where (P ◦ W ) is the element-wise product between P and W and where Md =
[δi,j mi,j (W )].
Proof. By its definition, the first passage time satisfies the recursive formula:




ωi,j ,
with probability p,ij ,
(2.7)
Ti,j =



ωi,k + Tk,j ,
with probability pi,k , k 6= j.
Therefore, the results follows from taking the expectation:
E[Ti,j ] = ωi,j pi,j +
X
pi,k (E[Tk,j ] + ωi,k ).
k6=j
The matrix M, which we call the mean first passage time matrix for a doublyweighted graph thus satisfies an equation similar to (2.1) for the passage time
matrix M of a graph with a single weight matrix, the transition matrix P . In fact,
we see that equation (2.6) is equivalent to (2.1) when wi,j = 1 for all (i, j) ∈ E
(i.e., for an unweighted graph).
The random variable tracking the time interval between consecutive visits to a
node has been referred to as the refresh time of that node [68] and mi,i (W ) is the
expected value of the refresh time for the random walk. We now obtain a relation
between π and the refresh times mi,i (W ).
30
Chapter 2. The Weighted Hitting Time of a Random Walk
Theorem 6 (Refresh times for doubly-weighted graphs). Consider a Markov
chain on a doubly-weighted graph G = (V, E, P, W ) with stationary distribution
π and associated mean first passage time matrix M. The refresh time for node i
is given by mi,i (W ) = (π T (P ◦ W )1n )/π i , implying that
Md = π T (P ◦ W )1n Md .
Proof. The stationary distribution of the transition matrix P is π ∈ Rn×1 . Therefore, premultiplying equation (2.6) by π T , we obtain
0 = π T (P ◦ W )1n 1Tn − π T Md ,
where the left hand side of equation (2.6) is zero since π T (I − P ) = π T − π T = 0.
Now we have that (π T (P ◦ W )1n )1Tn = π T Md . Since Md is a diagonal matrix
and π T (P ◦ W )1n is a scalar, we get that π i mi,i (W ) = π T (P ◦ W )1n . Thus,
dividing by π i we have that mi,i (W ) = π T (P ◦ W )1n /π i , and in matrix form
Md = π T (P ◦ W )1n diag({1/π 1 , . . . , 1 π n }) = π T (P ◦ W )1n Md .
This theorem implies that the refresh time mi,i (W ) of the random walk is
directly proportional to the visit frequencies 1/π i . Therefore, the relative visit
frequencies of one node compared to another are not a function of the weight
matrix W and only depend on the stationary distribution of the transition matrix
P.
31
Chapter 2. The Weighted Hitting Time of a Random Walk
We now investigate the properties of the hitting time of the weighted random
walk. The hitting time for a doubly-weighted graph G = (V, E, P, W ) with associated passage times matrix M is given by HW = π T hW , where hW = Mπ is the
vector of first passage times and the i−th entry hW,i in hW denotes the mean time
to go from i to any other node. We refer to the hitting time for a doubly-weighted
graph, HW , as the weighted hitting time. We now provide an analytic expression
for the vector hW ∈ Rn×1 .
Lemma 7. (First passage times for a doubly-weighted graph): Given a Markov
chain on a doubly-weighted graph G = (V, E, P, W ) with stationary distribution π
and associated first passage time matrix M, the following equality holds:
(I − P )hW = (P ◦ W )1n − 1n π T (P ◦ W )1n ,
where hW = Mπ.
Proof. Post multiplying equation (2.6) on both sides by π gives
(I − P )Mπ =(P ◦ W )1n 1Tn π − P Md π,
(I − P )hW =(P ◦ W )1n − P (π T (P ◦ W )1n )1n
=(P ◦ W )1n − 1n π T (P ◦ W )1n .
32
(2.8)
Chapter 2. The Weighted Hitting Time of a Random Walk
The right hand side of (2.8) gives the insight that, in general, hW,i 6= hW,j on
the doubly-weighted graph, unlike the counterpart for the single-weighted graph
(where hi = H for all i ∈ {1, . . . , n}). Interestingly enough however, there does
exist a relation between the weighted hitting time HW and the hitting time H as
is stated in the following theorem, whose proof is postponed to Section 8.
Theorem 8. (Weighted hitting time of a Markov chain): For the doubly-weighted
graph G = (V, E, P, W ), the weighted hitting time HW of the Markov chain is given
by
HW = π T (P ◦ W )1n H,
(2.9)
where H is the hitting time associated with the irreducible transition matrix P
with stationary distribution π .
Remark 9. The expected number of hops to go from one node to another in a
Markov chain with transition matrix P is its hitting time. The expected distance
travelled (and hence time taken) executing one hop is
P
i πi
P
j
pij ωi,j = π(P ◦
W )1n . Hence, it is consistent with intuition that the expected time to travel from
one node to another should be Hπ T (P ◦ W )1n as is formally shown in Section 8.
Given the above results, we are now ready to state another problem of interest.
Problem 3. (Optimizing the weighted hitting time of a reversible Markov chain):
Given the stationary distribution π and graph G with vertex set V, edge set E and
33
Chapter 2. The Weighted Hitting Time of a Random Walk
weight matrix W , determine the transition probabilities P = [pi,j ] solving:
minimize
π T (P ◦ W )1n
subject to
n
X
Tr (I − Π1/2 P Π−1/2 + qq T )−1
pi,j = 1, for each i ∈ {1, . . . , n}
j=1
π i pi,j = π j pj,i , for each (i, j) ∈ E
0 ≤ pi,j ≤ 1, for each (i, j) ∈ E
pi,j = 0, for each (i, j) ∈
/ E.
The following theorem establishes the convexity of the weighted hitting time
for transition matrices with fixed stationary distribution.
Theorem 10 (Convexity of Problem 3). Given the graph G with vertex set V,
edge set E and weight matrix W , let PG,π denote the set of matrices associated
with G that are irreducible reversible Markov chains with stationary distribution
π. Then, PG,π is a convex set and P 7→ π T (P ◦ W )1n H(P ) is a convex function
over PG,π .
Proof. Let S denote the set of symmetric positive definite matrices, for any stationary distribution π ∈ Rn>0 , denote the set SG,P,π := {I − Π1/2 P Π−1/2 +
qq T | P ∈ PG,π }. The proof of convexity of the set PG,π is similar to that of
the proof of Pπ in Theorem 4 and so is omitted for brevity. Then from the proof
of Theorem 4 we know there exists an affine mapping g(X) : PG,π 7→ SG,P,π given
34
Chapter 2. The Weighted Hitting Time of a Random Walk
by g(X) = I − Π1/2 P Π−1/2 + qq T . We know from [34] that f (X) = Tr[X −1 ] is
convex, therefore the perspective function h(X, t) = {tf (X/t) | t > 0} is also convex [10, Chapter 3.2.6]. Moreover the composition of h(X, t) with the affine mapping g(X), h(g(X), t) is also convex. Let t = (π T (X ◦ W )1n )1/2 , and notice that
π T (X ◦W )1n > 0 for X ∈ PG,π and therefore t > 0. Also notice that for a constant
k ∈ Rn>0 and matrix X ∈ Rn×n that Tr[( Xk )−1 ] = kTr[X −1 ]. Then h(g(X), t) =
tTr[( g(X)
)−1 ] = t2 Tr[(g(X)−1 ] = π T (X ◦ W )1n Tr[(I − Π1/2 XΠ−1/2 + qq T )−1 ] for
t
X ∈ PG,π .
2.3.2
SDP framework for optimizing the weighted hitting
time
In a similar fashion to Problem 1, we can formulate Problem 3 as an SDP by
introducing the symmetric slack matrix X ∈ Rn×n and the scalar variable t as is
shown in the following.
Problem 4. (Optimizing the weighted hitting time of a reversible Markov chain
(SDP)): Given the stationary distribution π and graph G with vertex set V, edge
35
Chapter 2. The Weighted Hitting Time of a Random Walk
set E and weight matrix W , determine Y = [yi,j ], X and t solving:
minimize Tr[X]
subject to


 t(I + qq T ) − Π1/2 Y Π−1/2 I 

0


I
X
n
X
yi,j = t, for each i ∈ {1, . . . , n}
j=1
π i yi,j = π j yj,i , for each (i, j) ∈ E
0 ≤ yi,j ≤ t, for each (i, j) ∈ E
yi,j = 0, for each (i, j) ∈
/E
π T (Y ◦ W )1n = 1
t ≥ 0.
Then, the transition matrix P is given by P = Y /t.
As in Problem 2, the first inequality constraint in Problem 4 represents an LMI.
Before noting when the LMI holds, first note that the constraints in Problem 4
imply that P t = Y and that t =
1
.
π T (P ◦W )1n
Hence, using a similar argument as
in Problem 2, the LMI constraint holds true if and only if X π T (P ◦ W )1n (I −
Π1/2 P Π−1/2 + qq T )−1 where X and π T (P ◦ W )1n (I − Π1/2 P Π−1/2 + qq T )−1 are
both positive definite, and therefore the SDP given by Problem 4 minimizes the
weighted hitting time.
36
Chapter 2. The Weighted Hitting Time of a Random Walk
2.3.3
Minimizing single hop distance
We now look at the objective of minimizing the mean time for a single hop of
a random walk. At time k, let Si,j be the time required to transition from i to j
in a single hop along an edge of length ωi,j . Then,
E [S] =
n X
n
X
pi,j Si,j
i=1 j=1
=
n
X
P [Xk = i]
i=1
=
n X
n
X
n
X
ωi,j P [Xk+1 = j]
j=1
π i ωi,j pi,j = π T (P ◦ W )1n .
(2.10)
i=1 j=1
The function π T (P ◦ W )1n is clearly convex in P . If one assumes that ωi,i = 0 for
all i ∈ {1, . . . , n}, then minimizing (2.10) over P yields the trivial solution P = I.
We can take into account both the single hop distance as well as the hitting time
to design a Markov chain as follows.
Problem 5. (Optimizing hitting time and mean distance): Given the stationary
distribution π and graph G with vertex set V, edge set E and weight matrix W ,
and given user specified constant α ∈ [0, 1], determine the transition probabilities
37
Chapter 2. The Weighted Hitting Time of a Random Walk
P = [pi,j ] solving:
minimize αTr (I − Π1/2 P Π−1/2 + qq T )−1
+ (1 − α)π T (P ◦ W )1n
subject to
n
X
pi,j = 1, for each i ∈ {1, . . . , n}
j=1
(2.11)
π i pi,j = π j pj,i , for each (i, j) ∈ E
0 ≤ pi,j ≤ 1, for each (i, j) ∈ E
pi,j = 0, for each (i, j) ∈
/ E.
This problem is convex since the sum of two convex problems is convex, moreover, it can be extended to an SDP utilizing the LMI defined in Problem 2. In
the context where ωi,i = 0 for all i ∈ {1, . . . , n}, varying the parameter α can be
used to control the connectivity of the Markov chain; the choice α = 1 ensures
connectivity, and the choice α = 0 minimizes the single hop distance while making
the graph disconnected.
2.4
Applications of the hitting time to surveillance
The results on hitting time for doubly-weighted graphs (i.e., the weighted
hitting time) presented in this work provide a general framework which can po-
38
Chapter 2. The Weighted Hitting Time of a Random Walk
tentially be applied to the analysis and design in a myriad of fields. We focus
on one application in particular; the intruder detection and surveillance problem.
We look at two variations of this problem:
Scenario I In practical stochastic intruder detection and surveillance scenarios, there is often a desire to surveil some regions more than others (i.e.,have a
pre- specified stationary distribution) while simultaneously minimizing the time
any one region has to wait before it is serviced. For this setup, in every step of
the random walk, the agent must move to a new region and execute its surveillance task. Assuming we are working on a doubly-weighted graph described by
G = (V, E, P, W ), let us also assume there is a fixed persistent intruder in the
environment and it takes si time for an agent to determine if the intruder is in
region i ∈ V. Denote the time to move from region i to region j by di,j , where
we can assume, without loss of generality, that di,i = 0. Then, we can define the
weight corresponding to the edge from i to j as ωi,j = di,j + sj . In this scenario
we wish to minimize the expected time to capture the persistent intruder when
no prior knowledge of their position is known.
Scenario II In this scenario we consider the intruder detection problem and
adopt a similar setup to Scenario I, however, we now assume a set of intruders
are distributed throughout the environment. Each intruder performs a malicious
39
Chapter 2. The Weighted Hitting Time of a Random Walk
activity in its host region for a fixed duration of time, which we call the intruder
life-time, followed instantaneously by another intruder. The intruder is caught
only if the agent is in the same region as the intruder for any duration of the
intruder life-time. For simplicity only a single intruder appears at a time.
In the following subsections we analyze the performance of various stochastic
surveillance policies as applied to Scenario I and Scenario II described above.
More specifically, we gauge the performance of other well-known Markov chain
design algorithms against the algorithms presented in this paper.
2.4.1
Optimal strategy for Scenario I
In the context of Scenario I the weighted hitting time of the agent’s transition matrix, P , corresponds to the average time it takes to capture an intruder
regardless of where the agent and intruder are in the environment. Therefore
by definition of the hitting time, we have the following immediate corollary for
Scenario I.
Corollary 11 (Optimal surveillance and service strategy). The transition matrix
P which has minimal hitting time is the optimal strategy for the intruder detection
problem described by Scenario I.
40
Chapter 2. The Weighted Hitting Time of a Random Walk
This tells us that if we restrict ourselves to reversible Markov chains, then not
only is the chain with minimal hitting time optimal, but given the results from
Section 2.2 and 2.3, we can also optimally design this chain.
2.4.2
Numerical analysis of Scenario II
In Scenario II the transition matrix with minimum hitting time is not guaranteed to be the optimal policy, and thus to gauge its performance compared to
other policies we analyze both homogeneous (uniform service/travel times) and
heterogeneous environment cases. To compare performance we generate a random
walk for the environment using the Metropolis-Hastings, fastest mixing Markov
chain (FMMC) [9], and hitting time algorithms. While game theoretic frameworks [19, 8] also generate stochastic policies, they are based on assumptions on
the intruder behavior. We avoid such assumptions here and, therefore, omit them
from our comparative analysis.
We first look at the homogeneous case which is described by the discretized
environment shown in Figure 2.2. We assume that a single surveillance agent executes a random walk in the environment, spending 1 time unit in each region, and
that the agent transitions between two connected regions instantaneously. Also,
we assume a uniform stationary distribution on the transition matrix (each node
in the region must be visited with equal likelihood). The Markov chain generated
41
Chapter 2. The Weighted Hitting Time of a Random Walk
by the Metropolis-Hastings algorithm is generated by applying the algorithm to
the random walk described by pi,j = 1/di for i 6= j and pi,j = 0 for i = j, where
di is the degree of a node i (excluding self-loops) [39]. The intruder life-time is
set to 66 time units and 500 intruders appear per simulation run (the sequence in
which the intruders appear is determined before each simulation run), for a total
simulation run of 33, 000 time units. As stated in the scenario description, the
intruder is caught if the surveillance agent is in the same region as the intruder for
any portion of the intruder life-time. Table 2.1 summarizes the statistical performance of each algorithm after 200 runs of the simulation and justifies our use of
the hitting time algorithm as a valid surveillance strategy; the hitting time algorithm captures intruders more frequently than the other two algorithms, and its
worst case performance is still better than the worst case performance of the other
two algorithms. Although results for an intruder life-time of only 66 time units
are presented here, we have found that the hitting time algorithm always outperforms the other two algorithms or is equivalent; the algorithms become equivalent
in the limiting case, when the intruder life-times are so low that no intruder can
be caught, or when the intruder-life times are so large that the intruder is always
caught.
For the heterogeneous case, we work with the environment shown in Figure 2.3.
In this environment the time taken by the agent to travel an edge is no longer
42
Chapter 2. The Weighted Hitting Time of a Random Walk
Figure 2.2: Environment with two obstacle represented by an unweighted graph.
Algorithm
Min
Hitting time
26.6
FMMC
24.6
Metropolis-Hastings 24.8
Mean
32.4
29.8
31.1
Max
38.2
34.4
37
StdDev
2.1
1.9
2.1
H
207
236
231
Table 2.1: Statistics on the percentage of intruders caught in 200 simulation runs
for the environment in Fig. 2.2.
instantaneous and has travel weights as shown in the figure. Once in a new region,
the agent is required to spend 1 time unit examining the region for malicious
activities. We again assume that each node in the region must be visited with equal
likelihood. We again also assume an intruder is caught if the surveillance agent is
in the same region as a intruder for any portion of the intruder life-time, but now
set the intruder life-time to 11 time units with a intruder appearing 500 times
(total of 5500 time units per simulation run). Since the design of the FMMC and
Metropolis-Hastings algorithms do not inherently account for non-uniform travel
and service times, we also compare the performance of the random walk generated
43
Chapter 2. The Weighted Hitting Time of a Random Walk
by the weighted hitting time algorithm against the random walk generated by
solving Problem 5 with α = 0.1 (smaller α emphasizes minimizing the length
of the average edge traveled in the graph). Table 2.2 summarizes the statistical
performance of each algorithm after 200 runs of the simulation. The weighted
hitting time algorithm’s performance compared to the FMMC and MetropolisHastings stochastic policies in this scenario is significantly better than what was
seen in the first scenario. We also note that the random walk policy determined by
solving Problem 5 performs comparably to the weighted hitting time policy. This
is to be expected since the Metropolis-Hastings and FMMC stochastic policies
do not account for heterogeneous travel/service times on the graph. To get a
better understanding of each algorithm’s performance in this intruder scenario,
the simulation is run for different intruder life-times, the results of which can be
seen in Figure 2.4. There are several key items worth noting from the simulation.
First, we see that the weighted hitting time algorithm significantly outperforms
the other algorithms for a large range of intruder life-times. This matches our
intuition since the algorithm inherently minimizes average travel time between
nodes. Second, notice that the Markov chain generated by solving Problem 5 (with
α = 0.1) performs well for small intruder life-times but its performance plateaus
quickly. This is due to the fact that the transition matrix generated by solving
Problem 5 forces agents to stay at a given node rather than jump nodes; as one
44
Chapter 2. The Weighted Hitting Time of a Random Walk
SEA
JFK
LAX
ANC
ORL
1
SEA
6
6
3
JFK
1
2
6
12
1
5
3
LAX
5
ANC
1
15
ORL
1
Figure 2.3: Various airport hub locations (top), and the corresponding weight
map (bottom). Edge weights between two hubs account for travel time between
hubs plus required service time once at hub. Self loops have no travel time so
encompass only service time required at hub.
would suspect, once intruder life-times increase, a strategy which places emphasis
on an agent that moves between regions will begin to perform relatively better.
Finally, observe that as intruder life-time increases, the algorithms’ capture rates
start to converge. As in the homogeneous case, this is due to the fact that once the
intruder’s life-time is long enough, the agent will almost surely reach the intruder
regardless of the policy it follows.
45
Chapter 2. The Weighted Hitting Time of a Random Walk
Algorithm
Weighted Hitting Time
Hitting Time+Mean Dist.
FMMC
Metropolis-Hastings
Min
44
40.6
29.8
30.4
Mean
50.1
47.1
35.4
36
Max
56
53
40.4
41.6
StdDev
2.2
2.2
2.2
2.1
HW
19.5
23.1
26.2
26.5
Table 2.2: Statistics on the percentage of intruders caught in 200 simulation runs
for the environment in Fig. 2.3.
1
0.9
% of captures
0.8
0.7
0.6
0.5
0.4
10
15
20
25
30
35
40
45
50
55
60
Intruder life−time
Figure 2.4: Percentage of intruders detected for varying intruder life-times by
a surveillance agent executing a random walk according to the Markov chain
generated by the hitting time algorithm (circle), FMMC algorithm (square), MH algorithm (asterisk), and the Markov chain generated by solving Problem 5
(diamond). Average points and standard deviation error bars are taken over 200
runs, where the intruder appears 500 times for each run.
To solve for the Markov chains with minimal hitting time (Problem 2 and
Problem 4) and with fastest mixing rate, we use CVX, a Matlab-based package
for convex programs [37]. The execution time to solve each Markov chain for the
46
Chapter 2. The Weighted Hitting Time of a Random Walk
examples described above takes on the order of a couple seconds using a computer
with a 1.3 GHz processor.
2.5
Summary
We have studied the problem of how to optimally design a Markov chain which
minimizes the hitting time to go from one region to any other region in a connected
environment. We have presented the first formulation of the hitting time for a
doubly-weighted graph, which we refer to as the weighted hitting time, and have
also provided a provably correct convex formulation for the minimization of both
the hitting time and the weighted hitting time. Finally, we have shown that
both problems can be written as SDPs and, moreover, have demonstrated the
effectiveness of using a Markov chain with minimal hitting time as a surveillance
policy as compared to other well-known Markov chain policies.
In the next chapter we look at extending the hitting time of a Markov chain
to multiple random walkers.
47
Chapter 2. The Weighted Hitting Time of a Random Walk
2.6
2.6.1
Proofs and supplemental material
Proof of Theorem 8
Proof of Theorem 8. Let β = π T (P ◦ W )1n , then from Theorem 6 we have that
Md from (2.6) can be written as βMd . Now from Theorem 13 the general solution
to (2.6) is
M = G((P ◦ W )1n 1Tn − βP Md ) + (I − G(I − P ))U,
(2.12)
where G is a generalized inverse of (I − P ) (see Theorem 15) and U is an arbitrary
matrix as long as the consistency condition
I − (I − P )G (P ◦ W )1n 1Tn − βP Md = 0
(2.13)
holds. Substituting (2.18) from Lemma 16 in for (I − P )G in (2.13) gives that
I − (I − P )G (P ◦ W )1n 1Tn − βP Md
tπ T
((P ◦ W )1n 1Tn − βP Md ),
T
π t
t
= T (π T (P ◦ W )1n 1Tn − βπ T P Md ),
π t
t
= T (β1Tn − β1Tn ) = 0,
π t
=
and so we have that the system of equations is consistent. This implies we can
design U to reduce (2.12). We start by seeing how the second term in (2.12)
can be reduced. Using (2.19) from Lemma 16 we have that (I − G(I − P ))U =
48
Chapter 2. The Weighted Hitting Time of a Random Walk
1n u T
U
u T 1n
= 1n kT , where kT =
uT U
.
u T 1n
Hence, we can re-write (2.12) as
M = G((P ◦ W )1n 1Tn − βP Md ) + 1n kT ,
(2.14)
designing U reduces to designing the n elements of k. Let K = diag[k], then
1n kT = 1n 1Tn K. Also, let Ξ = 1n π T , where Ξd = diag[π]. Utilizing these
expressions in (2.14) and taking the diagonal elements gives
M = G((P ◦ W )1n 1Tn − βP Md ) + 1n 1Tn K d ,
=⇒ βMd = (G(P ◦ W )Ξ)d Md − β(GP )d Md + K,
=⇒ K = βMd − (G(P ◦ W )Ξ)d Md + β(GP )d Md ,
where we use Lemma 14 to get the initial diagonal reduction. Substituting the
expression for K into (2.14), and recalling that 1n kT = 1n 1Tn K gives
M = G(P ◦ W )Ξ − 1n 1Tn (G(P ◦ W )Ξ)d
(2.15)
+ β(1n 1Tn (GP )d − GP + 1n 1Tn ) Md ,
where we use the fact that 1n 1Tn = ΞMd . Now from (2.19) we have that I − G −
GP =
1n u T
.
uT 1n
T
Notice that 1n 1Tn (I − G − GP )d = 1n 1Tn ( 1unTu1n )d =
1n uT
uT 1n
and so this
implies that 1n 1Tn − 1n 1Tn Gd + 1n 1Tn (GP )d = I − G + GP , which implies that
1n 1Tn + 1n 1Tn (GP )d − GP = I − G + 1n 1Tn Gd . Substituting this into (2.15) gives
49
Chapter 2. The Weighted Hitting Time of a Random Walk
the following reduced form.
M = G(P ◦ W )Ξ − 1n 1Tn (G(P ◦ W )Ξ)d
(2.16)
+ β(1n 1Tn Gd + I − G) Md .
Observing the definition of the generalized inverse, G, given by Theorem 15 part
(ii) and recalling that Ξ = 1n π T , we can rewrite the first term on the right hand
side of (16) as G(P ◦ W )Ξ = (I − P + tuT )−1 (P ◦ W )1n π T . Substituting (2.20)
in for the right hand side with t = (P ◦ W )1n gives G(P ◦ W )Ξ =
and so 1n 1Tn (G(P ◦ W )Ξ)d = 1n 1Tn ( uT11n Ξ)d =
1
Ξ
uT 1n
1n π T
uT 1n
=
1
Ξ,
uT 1n
= G(P ◦ W )Ξ. Therefore,
the first two terms in (2.16) cancel giving the equality
M = β(1n 1Tn Gd + I − G)Md .
(2.17)
We have already defined t in the generalized inverse G but not u. Let u = π and
multiply the right hand side of (2.17) by π and the left hand side by π T . Utilizing
equality (2.21) from Lemma 16 gives
π T Mπ = π T β(1n 1Tn Gd + I − G)Md π
= β(1Tn Gd + π T −
πT
)1n
β
= β(1Tn Gd 1n + 1 −
1
)
β
= β(Tr[G] + 1) − 1.
50
Chapter 2. The Weighted Hitting Time of a Random Walk
Noting that the eigenvalue at 1 for an irreducible row-stochastic matrix is unique,
it can be easily verified using the orthogonality property of left and right eigen¯ i = (1 − λi ) for i ∈ {2, . . . , n}, where λi
vectors that the eigenvalues of G−1 are λ
¯ 1 . Taking the
are eigenvalues of P and λi 6= 1. Therefore, it only remains to find λ
trace of G−1 gives Tr[I − P + (P ◦ W )1n π T ] = Tr[I − P ] + Tr[(P ◦ W )1n π T ] =
Pn
i=1 (1−λi )+π
T
¯ 1 = π T (P ◦W )1n = β. Therefore,
(P ◦W )1n , which implies that λ
β(Tr[G] + 1) − 1 = β( β1 +
2.6.2
Pn
1
i=2 1−λi
+ 1) − 1 = β(1 +
Pn
1
i=2 1−λi ).
Supplemental Material
For completeness, we include the following results which are needed in the
proof of Theorem 8. We begin with some standard results on generalized inverses.
For more details refer to [43, Chapter 4] or [42] .
Definition 12 (Generalized inverse). A generalized inverse of an m × n matrix
A is defined as any n × m matrix A− that has the property
AA− A = A.
It should be noted that a generalized inverse always exists, although it is not always unique. However, for non-singular matrices the generalized inverse is unique
and corresponds to the usual notion of a matrix inverse. The following theorems
summarize practical considerations when working with generalized inverses.
51
Chapter 2. The Weighted Hitting Time of a Random Walk
Theorem 13. The equation Ax = b admits a solution if and only if every generalized inverse A− satisfies
AA− b = b.
Then, we say Ax = b is consistent and all its general solutions are given by
x = A− b + (A− A − I)z,
where z is an arbitrary vector. Moreover, a necessary and sufficient condition for
the equation AX = C to be consistent is that (I − AA− )C = 0, in which case the
general solution is given by
X = A− C + (I − A− A)U,
where U is an arbitrary matrix.
The next two results come from [44, Chapter 7].
Lemma 14 (Diagonal matrix properties). For π with positive non-zero elements,
let 1n π T = Ξ, where Ξd = diag[π]. Also, let Λ be any diagonal matrix, X any
square matrix of same dimensions as Λ, and D = (Ξd )−1 , then
(i.) (XΛ)d = (Xd )Λ, and
(ii.) (X1n 1Tn )d = (XΞ)d D, and
52
Chapter 2. The Weighted Hitting Time of a Random Walk
(iii.) 1n 1Tn Ξd = Ξ.
Theorem 15 (Generalized inverse of I−P ). Let P ∈ Rn×n be the transition matrix
of a irreducible Markov Chain with stationary distribution π. Let u, t ∈ Rn be
any vectors such that uT 1n 6= 0 and π T t 6= 0, then
(i.) I − P + tuT is nonsingular, and
(ii.) (I − P + tuT )−1 is a generalized inverse of I − P .
Lemma 16 (Properties of the generalized inverse of I − P ). Let G = (V, E, P, W )
be a doubly-weighted graph with associated weight matrix W and irreducible transition matrix P with stationary distribution π. Also let G = (I − P + tuT )−1
denote the generalized inverse of (I − P ), then the following relations hold.
tπ T
,
πT t
(2.18)
1n uT
, and
uT 1n
(2.19)
(I − P )G = I −
G(I − P ) = I −
1n
= Gt.
uT 1n
(2.20)
If t = (P ◦ W )1n and uT = π T then
πT G =
πT
π T (P ◦ W )1n
53
(2.21)
Chapter 2. The Weighted Hitting Time of a Random Walk
Proof. First, notice that (I − P + tuT )(I − P + tuT )−1 = I implies that
(I − P )G = I − tuT G.
(2.22)
Multiplying both sides on the left by π T and noting that π T (I − P ) = 0 gives
that π T = (π T t)uT G. Dividing through by (π T t) gives
πT
= uT G,
T
π t
(2.23)
and substituting (2.23) into (2.22) gives (2.18).
Following a similar procedure as before with (I −P +tuT )−1 (I −P +tuT ) = I,
where we now multiply both sides on the right by 1n and note that (I − P )1n = 0
results in (2.20), which after appropriate substitution gives (2.19).
For the proof of equality (2.21), first we check that t = (P ◦W )1n and uT = π T
satisfy the conditions of Theorem 15. The definition of W guarantees that P ◦ W
has at least one non-zero element which implies π T t = π T (P ◦ W )1n 6= 0. Also,
uT 1n = π T 1n = 1. Now substituting u and t into (2.23) gives (2.21).
54
Chapter 3
The Hitting Time of Multiple
Random Walks
This chapter further extends notions of the hitting time described in Chapter 2
and is organized as follows. In Section 3.1 we introduce special tensor notation that
is specific to this chapter only and review useful concepts of Kronecker products
and Markov Chains. In Section 3.2 we provide an alternate formulation for the
hitting time and weighted hitting time. In Section 3.3 we introduce the ‘group’
hitting time and hitting time between sets of nodes for a Markov chain, and
provide a detailed characterization. In Section 3.4 we provide insight into optimal
group hitting times through numerical simulation, and finally in Section 3.5 we
summarize our findings.
As in the previous chapter, we work with finite irreducible time-homogeneous
Markov chains (see Section 2.1 for details).
55
Chapter 3. The Hitting Time of Multiple Random Walks
3.1
Tensor notation and the Kronecker product
In this section we define various useful concepts and notation. In particular,
we introduce the notation that will be used throughout this chapter to deal with
tensors, and conclude with a brief summary of the Kronecker product (also known
as a tensor product) and some of its properties.
We use the notation A = [ai1 ...ik ,j ] to denote the matrix generated by elements
ai1 ...ik ,j , where the j index denotes the j-th column of A and the rows of A are
determined by cycling through index ik , then ik−1 , and so forth. For example,
consider i1 , i2 , j ∈ {1, . . . , n}, then














A = [ai1 i2 ,j ] = 














a11,1 a11,2 . . .
a12,1 a12,2 . . .
..
.
..
.
...
a1n,1 a1n,2 . . .
a21,1 a21,2 . . .
..
.
..
.
...
..
.
..
.
...
ann,1
...
...
a11,n 


a12,n 


.. 
. 


.. 
. 

.
.. 
. 


.. 

. 



... 


ann,n
For the case where A = [ai,j ] this corresponds to the classic interpretation with
element ai,j in the i-th row and j-th column of A. To avoid ambiguity, especially
in cases when A = [ai1 ...ik ,j ], at times we will refer to the (i, j) element of A by
56
Chapter 3. The Hitting Time of Multiple Random Walks
A(i, j). Unless otherwise mentioned, vectors will be denoted by bold-faced letters
(i.e., a). We use the notation diaga to indicate the diagonal matrix generated by
vector a and vec[A] to indicate the vectorization of a matrix A ∈ Rn×m where
vec[A] = [A(1, 1), . . . , A(n, 1), A(1, 2), . . . , A(n, 2), . . . , A(m, 1), . . . , A(n, m)]T .
In other words, even if we define A as A = [ai1 ...ik ,j ], the vector vec[A] = vec[[ai1 ...ik ,j ]]
is simply a stacking of the columns of A. We also define the special matrix
...hn ,k
[Iih11...i
] as the matrix whose entries are all zero except for a single entry at
n ,j
(h1 . . . hn , k) which has a value of 1, where h1 , . . . , hn , k can only take values within
the range of values that i1 , . . . , in , j take. With this matrix definition, it is easy
...hn ,k T
to verify for A = [ai1 ...in ,j ] that ah1 ...hn ,k = vec[[Iih11...i
]] vec[A]. This enables
n ,j
us to go back and forth between the vectorized notation to the individual matrix
elements. We denote In ∈ Rn×n as the identity matrix of size n, 1n as the vector
of ones of size n and 0n×n as the matrix zeros of size n×n. We define a generalized
Kronecker delta function δi1 i2 ...in ,j , by




1, if ik = j for any k ∈ {1, . . . , n},
δi1 ...in ,j =



0, otherwise .
Then, with a slight abuse of notation, Ad = [δi1 i2···n ,j ai1 ...ik ,j ] represents the “diagonal” matrix generated by the elements of A. In reality, only when A = [ai,j ] is
the matrix truly diagonal. Finally, since the subscript operator is already in use,
57
Chapter 3. The Hitting Time of Multiple Random Walks
we use the superscript operation in parenthesis to delineate between two variables
with the same name. For example we write A(1) and A(2) to distinguish that the
matrices are different. A superscript without parenthesis denotes a matrix raised
to that power (i.e., A2 = AA and (A(2) )2 = A(2) A(2) ).
We are now ready to review some useful facts about Kronecker products. The
Kronecker product, represented by the symbol ⊗, of two matrices A ∈ Rn×m and
B ∈ Rq×r is a nq × mr matrix given by

 a1,1 B . . . a1,m B

 .
..
..
A⊗B = 
.
.
 ..


.
an,1 B . . an,m B




.



To build some intuition, notice for A ∈ Rn×n , that In ⊗ A is the block diagonal
matrix with n copies of A on the diagonal:

 A 0n×n . . . 0n×n


..
..
..
 0
.
.
.
 n×n
In ⊗ A = 
 .
...
...
 ..
0n×n



0n×n . . . 0n×n
A






.





(3.1)
This implies for A = In , that In ⊗ A = In ⊗ In = In2 . The Kronecker product
is bilinear and has many useful properties, two of which are summarized in the
following lemma; see [41] for more information.
58
Chapter 3. The Hitting Time of Multiple Random Walks
Lemma 17. Given the matrices A, B, C and D, the following relations hold for
the Kronecker product.
(i) (A ⊗ B)(C ⊗ D) = (AC) ⊗(BD),
(ii) (B T ⊗ A) vec[C] = vec[ACB],
where it is assumed that the matrices are of appropriate dimension when matrix
multiplication or addition occurs. In addition, for matrices A ∈ Rn×n and B ∈
B
Rm×m with respective eigenvalues λA
i , i ∈ {1, . . . , n} and λj , j ∈ {1, . . . , m},
B
(iii) the eigenvalues of A ⊗ B are λA
i λj for i ∈ {1, . . . , n} and j ∈ {1, . . . , m}.
3.2
The hitting time of a Markov chain
We begin by recalling some properties of the single random walker hitting time
from Chapter 2, and then provide a new formulation for determining this quantity.
We show that this alternate formulation gives rise to new results for the pairwise
hitting time between two specified nodes, and, in the subsequent section, utilize it
to extend the notion of the hitting time and pairwise hitting time to the multiple
random-walker case.
59
Chapter 3. The Hitting Time of Multiple Random Walks
3.2.1
The hitting time of a Markov chain
Consider a connected undirected weighted graph G = (V, E, P ) with node set
V := {1, . . . , n}, edge set E ⊂ V × V, and weight matrix P = [pi,j ] with the
property that pi,j ≥ 0 if (i, j) ∈ E and pi,j = 0 otherwise. We interpret the
weight of edge (i, j) as the probability of moving along that edge. Therefore, P
is a transition matrix of a Markov chain, where the element pi,j in the matrix
represents the probability with which the random walk visits node j from node i.
As done previously, let Xt ∈ {1, . . . , n} denote the location of a random walker
at time t ∈ {0, 1, 2, . . . }. For any two nodes i, j ∈ {1, . . . , n}, the first passage
time from i to j, denoted by Ti,j , is the first time that the random walker starting
at node i at time 0 reaches node j, that is,
Ti,j = min{t ≥ 1 | Xt = j given that X0 = i}.
The mean first passage time from i to j is then given by mi,j = E[Ti,j ]. Denoting
the mean first passage time matrix M as the matrix whose i, jth entries are given
by mi,j , recall that the hitting time, H(P ), can be written as H(P ) = π T M π.
Given the definition of the hitting time, one quickly sees that it can also be
determined using the matrix M as follows,
π T M π = (π ⊗ π)T vec[M ] =
n
X
i=1
60
πi
n
X
j=1
π j mi,j = H,
(3.2)
Chapter 3. The Hitting Time of Multiple Random Walks
where the relation π T M π = (π ⊗ π)T vec[M ] follows from identity (ii) in Lemma
17. From Chapter 2 we saw that mi,j is described by the following recursive
formula.
mi,j = pi,j +
n
X
n
X
pi,k (mk,j + 1) = 1 +
k=1,k6=j
pi,k mk,j .
k=1,k6=j
The above formula can be expressed in various vectorized forms. The classical
form already seen utilizes the matrix representation of M and is given by
(I − P )M = 1n 1Tn − P Md ,
(3.3)
where we recall that Md = diag{1/π 1 , . . . , 1/π n } and the value for Md is determined by pre-multiplying (3.3) by π T . For reasons which will become clear
later, we vectorize the matrix M to generate a different representation of (3.3).
Applying property (ii) of Lemma 17 to (3.3) gives
(In ⊗(In − P )) vec[M ] = 1n2 − (In ⊗ P ) vec[Md ].
(3.4)
Similar to before, Md can be determined from (3.4) with appropriate vector premultiplication.
We know from Theorem 1 that the hitting time can be written as a function
of the eigenvalues of P as
H(P ) = 1 +
n
X
i=2
1
.
1 − λi
The proof for Theorem 1 relies on the fact that Md is known, however, as will be
seen next, there exists an equivalent expression for H(P ) which requires no such
61
Chapter 3. The Hitting Time of Multiple Random Walks
knowledge. Before presenting the alternate formulation for H(P ), we introduce
the following useful result.
Lemma 18. (Eigenvalue shifting for stochastic matrices): Let P ∈ Rn×n be an
irreducible row stochastic matrix, and let E be any diagonal matrix with diagonal
elements Eii ∈ {0, 1}, with at least one diagonal element which is zero. Then the
eigenvalues λi of P E satisfy |λi | < 1 for all i ∈ {1, . . . , n}.
Proof. Consider the case when the kth diagonal element of E is equal to zero,
Ekk = 0 and all other diagonal elements are equal to 1, Eii = 1 for i 6= k.
This has the affect of making the kth column of P zero, pik = 0; the physical
interpretation is that no node can move to node k. Thus the matrix A = P E is
reducible and there exists a permutation matrix S ∈ Rn×n such that SAS T is the
block upper triangular matrix


(1)

SAS = 

T
Br×r
Cr×(n−r)
0(n−r)×r
(2)
B(n−r)×(n−r)

,

(3.5)
where each B (i) is square. Clearly, SAS T remains sub-stochastic under permutation and the eigenvalues of the matrix A correspond to the eigenvalues of each
block B (i) . Consider the trivial permutation which moves the column k to the
first column. Then, B (1) = [0] ∈ R1×1 and B (2) ∈ Rn−1×n−1 . Clearly B (2) is
sub-stochastic since P is irreducible and there exists a path from every node to
62
Chapter 3. The Hitting Time of Multiple Random Walks
every other node. Now, if B (2) is also irreducible then λmax [B (2) ] < 1, [60]. If
it is not, then there exists another permutation matrix,S¯ ∈ Rn−1×n−1 , such that
¯ (2) S¯T is upper block triangular similar to (3.5). Due to the irreducibility of
SB
¯ (2) S¯T are sub-stochastic. Therefore, if they
P each of the diagonal blocks of SB
are irreducible then λmax < 1 for each block. If not, then we repeat the previous
argument until we are left with diagonal blocks which are irreducible or zero. The
case when more than one column is zero follows from noting that if ai,j ≤ bi,j for
each i and j then ρ[A] ≤ ρ[B], [60, Chapter 7.10].
We are almost ready to present our alternate representation of the hitting
time, but first we must introduce the following equality: notice that vec[Md ] =
E vec[M ], when E is defined by E = diag[δi,j ]. Using this interpretation of vec[Md ]
we are ready to state our result.
Theorem 19. (Hitting times of an irreducible Markov chain): Consider a Markov
chain with an irreducible transition matrix P ∈ Rn×n , then the following properties
hold:
(i) the hitting time of the Markov chain is given by
H(P ) = (π ⊗ π)T vec[M ],
where
vec[M ] = (In2 − (In ⊗ P )(In2 − E))−1 1n2 ,
63
and
Chapter 3. The Hitting Time of Multiple Random Walks
(ii) the pairwise hitting time between nodes h and k, denoted mh,k , of the Markov
chain is given by
h,k T
mh,k = vec[[Ii,j
]] vec[M ].
Proof. First notice that by rearranging (3.4) and substituting E vec[M ] for vec[Md ]
gives that
(In2 − (In ⊗ P )(In2 − E)) vec[M ] = 1n2 .
(3.6)
From (3.2) we know that H = (π ⊗ π) vec[M ], therefore it only remains to show
that (In2 −In ⊗ P (In2 −E)) is in fact invertible. First, recall from (3.1) that In ⊗ P
results in the block diagonal matrix, whose diagonal blocks consist of copies of P .
Second, notice that (In2 − E) is simply the identity matrix with some diagonal
entries set to zero. It can be easily verified that (In ⊗ P )(In2 − E) results in the
block diagonal matrix where each block contains the matrix P with one column
64
Chapter 3. The Hitting Time of Multiple Random Walks
set to zero. For example, for P ∈ R3×3 we
  0 p12 p13    0 p
p
22
23 
   0 p
p
32
33  


0 0 0



(In ⊗ P )(In2 − E) = 
0 0 0




0 0 0




0 0 0




0 0 0



0 0 0
have that

0 0 0
0 0 0
0 0 0
0 0 0
0
p
11
p
21
p
31
0 0
0 p13
0 p23
0 p33
0 0 0
0 0 0
0 0 0
0 0 0
0 0 0
0 0 0
0
p
11
p
21
p
31
0 0
p12
p22
p32
0 0 0 














.















Notice that each diagonal block will have at least one column set to zero. Hence,
using Lemma 18, we have that maximum eigenvalue of each block is strictly less
than one in magnitude, and thus λmax [(In ⊗ P )(In2 − E)] < 1. Let λi denote the
eigenvalues of (In ⊗ P )(In2 −E), then since the eigenvalues of (In2 −(In ⊗ P )(In2 −
E)) are simply 1 − λi and |λi | < 1 for all i ∈ {1, . . . , n2 }, this implies the matrix
(In2 − (In ⊗ P )(In2 − E)) is invertible and vec[M ] is the unique solution to (3.6).
Remark 20. Notice that when the transition matrix P is reducible, the hitting
time H(P ) is infinite (i.e., due to In2 −(In ⊗ P )(In2 −E) being singular). In other
words, the hitting time gives a notion of the connectivity of the graph.
65
Chapter 3. The Hitting Time of Multiple Random Walks
3.2.2
Hitting time for doubly-weighted graphs
Leveraging results from Chapter 2 , Theorem 19 can be easily extended to
account for “travel distances” between nodes. Define the doubly-weighted graph
G = (V, E, P, W ), where W = [ωi,j ] is the weight matrix with the properties that:
if (i, i) ∈ E, then ωi,i ≥ 0; if (i, j) ∈ E, i 6= j, then ωi,j > 0 ; and if (i, j) ∈
/ E,
then ωi,j = 0. Let P = [pi,j ] be the transition matrix associated with G with the
property that pi,j ≥ 0 if (i, j) ∈ E and pi,j = 0 otherwise.
Recall that the first passage time is now denoted by
Ti,j (W ) = min
k−1
nX
o
wXn ,Xn+1 , for k ≥ 1 | Xk = j given that X0 = i .
n=0
Letting mi,j (W ) = E[Ti,j ] denote the mean first passage time from i to j, or in matrix notation M = [mi,j (W )], the following matrix relationship can be determined
(see Section 2.3.1 equation (2.6)).
(I − P )M = (P ◦ W )1n 1Tn − P Md ,
(3.7)
where P ◦W denotes the Hadamard product between P and W . Similar to before,
applying property (ii) of Lemma 17 to (3.7) gives
(In ⊗(In − P )) vec[M] = vec[(P ◦ W )1Tn 1n ] − (In ⊗ P ) vec[Md ].
Given the above expression, the following Lemma follows from Theorem 19.
66
Chapter 3. The Hitting Time of Multiple Random Walks
Lemma 21. (hitting time and pairwise hitting times of a doubly-weighted graph):
Given the doubly-weighted graph G = (V, E, P, W ), the following properties hold:
(i) the weighted hitting time HW (P ) of the Markov chain is given by
HW (P ) = (π ⊗ π)T vec[M],
where
vec[M] = (In2 − (In ⊗ P )(In2 − E))−1 vec[(P ◦ W )1Tn 1n ], and
(ii) the pairwise hitting time between nodes h and k, denoted mh,k (W ), of the
Markov chain is given by
h,k T
mh,k (W ) = vec[[Ii,j
]] vec[M].
Now that the single random walker case has been explored in detail, we explore
the case of multiple random-walkers in the following section.
3.3
Group hitting time of multiple Markov chains
In the following sections, we will expand on the single agent hitting time to
the N -agent group hitting time. To build intuition, we initially assume that every
agent can reach all nodes in the graph, and then move to the case where each
agent only needs to have access to a subset of nodes in the graph.
67
Chapter 3. The Hitting Time of Multiple Random Walks
3.3.1
Random-walkers covering the full graph
Consider h ∈ {1, . . . , N } connected undirected weighted graphs G (h) = (V, E (h) , P (h) )
with same node sets V := {1, . . . , n} and different edge sets E (h) ⊂ V × V with
(h)
corresponding transition matrices P (h) = [pi,j ] for h ∈ {1, . . . , N } satisfying the
(h)
(h)
property pi,j ≥ 0 if (i, j) ∈ E (h) and pi,j = 0 otherwise. As before, each matrix
P (h) describes a Markov chain on the graph.
(1)
(2)
(N )
Let Xt , Xt , . . . , Xt
∈ {1, . . . , n} denote the location of N random walkers
at time t ∈ {0, 1, 2, . . . }. For any N + 1 nodes i1 , . . . , iN , j ∈ {1, . . . , n}, the
first passage time from any ih , h ∈ {1, . . . , N } to j, denoted by Ti1 ...iN ,j , is the
first time that any random walker reaches node j, when starting from nodes ih ,
h ∈ {1, . . . , N }. More formally,
(h)
Ti1 ...iN ,j = min{t ≥ 1 | For h ∈ {1, . . . , N }, any Xt
(h)
= j given that X0
= ih }.
(3.8)
With this definition, we are ready to state our first result for the N -random-walker
system.
Lemma 22. (Recursive formulation of first passage time for multiple randomwalkers): Let mi1 ...iN ,j = E[Ti1 ...iN ,j ] denote the first passage time from any ih ,
h ∈ {1, . . . , N } to j. Also, let P (h) be the transition matrix associated with G (h) ,
68
Chapter 3. The Hitting Time of Multiple Random Walks
then
X
mi1 ...iN ,j = 1 +
(1)
(N )
mk1 ...kN ,j pi1 ,k1 . . . piN ,kN ,
k1 ,...,kN 6=j
or, in matrix notation,
M = 1nN 1Tn + (P (1) ⊗ · · · ⊗ P (N ) )M − (P (1) ⊗ · · · ⊗ P (N ) )Md ,
(3.9)
where M = [mi1 ...iN ,j ] and Md = [δi1 ...iN ,j mi1 ...iN ,j ].
Proof. For clarity, we first study the 2-agent case and then generalize. By definition, the 2-agent first passage time satisfies the recursive formula



(1)
(2)
(1) (2)

1,
with probability pi1 ,j + pi2 ,j − pi1 ,j pi2 ,j ,
Ti1 i2 ,j =



Tk1 k2 ,j + 1, with probability p(1) p(2) such that k1 , k2 6= j.
i1 ,k1 i2 ,k2
Taking the expectation we have that
(1)
(2)
(1)
(2)
E[Ti1 i2 ,j ] = pi1 ,j + pi2 ,j − pi1 ,j pi2 ,j +
X
(1)
(2)
(E[Tk1 k2 ,j ] + 1)pi1 ,k1 pi2 ,k2
k1 ,k2 6=j
(1)
(2)
(1)
(2)
= pi1 ,j + pi2 ,j − pi1 ,j pi2 ,j +
X
k1 ,k2 6=j
(1)
(2)
mk1 k2 ,j pi1 ,k1 pi2 ,k2 +
X
(1)
(2)
pi1 ,k1 pi2 ,k2 .
k1 ,k2 6=j
(3.10)
69
Chapter 3. The Hitting Time of Multiple Random Walks
(1)
(2)
Utilizing the row-stochastic property of P (1) and P (2) expand (−pi1 ,j pi2 ,j ) above
to get
(1)
(2)
−pi1 ,j pi2 ,j = −(1 −
X
(1)
pi1 ,k1 )(1 −
k1 6=j
= −1 +
X
X
(2)
pi2 ,k2 )
k2 6=j
(1)
pi1 ,k1 +
k1 6=j
= −1 +
X
X
(2)
pi2 ,k2 − (
k2 6=j
(1)
pi1 ,k1 +
k1 6=j
X
X
(1)
pi1 ,k1 )(
k1 6=j
X
(2)
pi2 ,k2 −
k2 6=j
X
(2)
pi2 ,k2 )
k2 6=j
(1)
(2)
pi1 ,k1 pi2 ,k2 ,
k1 ,k2 6=j
substituting this back into (3.10) gives the result
mi1 i2 ,j = 1 +
X
(1)
(2)
mk1 k2 ,j pi1 ,k1 pi2 ,k2 ,
k1 ,k2 6=j
or, in matrix notation,
M = 1n2 1T2 + (P (1) ⊗ P (2) )M − (P (1) ⊗ P (2) )Md ,
where M = [mi1 i2 ,j ] and Md = [δi1 i2 ,j mi1 i2 ,j ].
Similar to the 2-agent case, the N -agent first passage time satisfies the recursive
formula
Ti1 ...iN ,j =




1,
(1)
(2)
(N )
with probability 1 − (1 − pi1 ,j )(1 − pi2 ,j ) . . . (1 − piN ,j ),



Tk1 ...kN ,j + 1, with probability p(1) p(2) . . . p(N ) such that k1 , . . . , kN 6= j.
i1 ,k1 i2 ,k2
iN ,kN
Similar to before, we take expectations and utilize the row-stochastic properties
of each P (i) to reach the result.
70
Chapter 3. The Hitting Time of Multiple Random Walks
Given the formulation for the N -agent first passage time matrix we can define
a quantity similar to the mean first passage time given by (3.2). In order to
do this, we first need to define the frequency of being at any given node in the
graph. This quantity should take into account the probability of being at one node
instead of another in the limit of the random walks. Since the random walks are
evolving in parallel, the relative frequency of being at a specific node is simply the
average of the N random walkers visit frequency at that node. More explicitly,
π ave =
PN
h=1 (π
(h)
)/N . Then, similar to the single agent case, the N -agent mean
first passage time from start nodes ih , h ∈ {1, . . . , N }, denoted hi1 ...iN , is given by
hi1 ...iN =
n
X
mi1 ...iN ,j π ave,j .
j=1
Therefore, the average time to go from any set of N nodes to a single node in a
graph is given by
HN =
n
X
i1 =1
(1)
π i1
···
n
X
(N )
π iN
iN =1
n
X
mi1 ...iN ,j π ave,j
j=1
= (π (1) ⊗ · · · ⊗ π (N ) )M π ave
= (π ave ⊗ π (1) ⊗ · · · ⊗ π (N ) ) vec[M ],
where we denote HN as being the N -agent group hitting time. It is clear the group
hitting time can be written as the function P (1) ×· · ·×P (N ) 7→ HN (P (1) , . . . , P (N ) ),
but to ease notation we simply write HN .
Given the definition of HN we are now ready to state our next result.
71
Chapter 3. The Hitting Time of Multiple Random Walks
Theorem 23. (Group hitting time for irreducible Markov chains): Consider N
multiple Markov chains, each with an irreducible transition matrix P (h) ∈ Rn×n .
P
N +1
N +1
(h)
Let π ave = ( N
)/N , and let E ∈ Rn ×n
be the diagonal matrix which
h=1 π
satisfies the equality E vec[M ] = vec[Md ]. Then the following properties hold:
(i) the group hitting time of the Markov chain is given by
HN = (π ave ⊗ π (1) ⊗ · · · ⊗ π (N ) )T vec[M ],
where
vec[M ] = (InN +1 − (In ⊗ P (1) ⊗ · · · ⊗ P (N ) )(InN +1 − E))−1 1nN +1 , and
(3.11)
(ii) the hitting time between nodes h1 , . . . , hN and k, denoted mh1 ...hN ,k , of the
Markov chain is given by
...hN ,k T
]] vec[M ].
mh1 ...hN ,k = vec[[Iih11...i
N ,j
Proof. Letting P = P (1) ⊗ · · · ⊗ P (N ) , notice that (3.9) can be written in the
vectorized form
vec[M ] = 1nN +1 + (In ⊗ P ) vec[M ] − (In ⊗ P ) vec[Md ].
Rearranging terms and substituting E vec[M ] for vec[Md ] gives
(InN +1 − (In ⊗ P )(InN +1 − E)) vec[M ] = 1nN +1 .
72
(3.12)
Chapter 3. The Hitting Time of Multiple Random Walks
To ease the complexity of our notation, we move forward by looking at the
2-agent case and then generalizing from there. For the two agent case with P =
P (1) ⊗ P (2) the system (3.12) becomes
(In3 − (In ⊗ P )(In3 − E)) vec[M ] = 1n3 ,
indicating a unique solution exists if (In3 − (In ⊗ P )(In3 − E)) is invertible. Let
(1)
the eigenvalues λi
(2)
and λi
be associated with transition matrices P (1) and P (2) ,
respectively. Then, from property (iii) of Lemma 17 the eigenvalues of P (1) ⊗ P (2)
(1) (2)
are λj λk
for j, k ∈ {1, . . . , n}. This means, when each P (i) is periodic, the
Kronecker product can results in a Markov chain which has multiple eigenvalues
at 1, making this chain reducible. Therefore, we must show that irreducible blocks
can be constructed that allow the application of Lemma 18 as before. As before
we require that (In ⊗ P )(In3 − E) has λmax < 1. Recall that In ⊗ P is a block
diagonal matrix consisting of copies of P . In each block, different columns of P
are set to zero when multiplied by (In3 − E), however, unlike the single agent case
(Theorem 19), now multiple columns are set to zero in each block by definition of
E. Therefore, we need only show that in every block, each irreducible component
of P has at least one column set to zero when multiplied by (In3 − E). Thus, we
first examine the structure of P . Since P = P (1) ⊗ P (2) is reducible we can apply
a series of permutation matrices S (i) for i ∈ {1, . . . , m} to P ([60, Chapter 8.3])
73
Chapter 3. The Hitting Time of Multiple Random Walks
such that


(1)
∗ ...
∗
 A


..
 0 A(2) . . .
.

P¯ = 
 .
... ...
 ..
∗



0
...
0 A(k)





.





(3.13)
Where P¯ = SP S T with permutations S = (S (m) ) . . . (S (1) ), and each A(i) is irreducible. Since P is row-stochastic then each irreducible component is either
sub-stochastic (ρ[A(i) ] < 1) or stochastic (ρ[A(i) ] = 1). Since we need only worry
about the stochastic A(i) , let’s denote B ∈ Rr×r as a irreducible stochastic matrix
A(i) ∈ Rr×r in (3.13). For B irreducible, P¯ has the form

(1)
∗ ... ... ...
∗
 A


..
... ... ... ...
 0
.


 .
...
 ..
B
0 ...
0

P¯ = 
 .
.. ..
 ..
.
. A(l) . . .
∗


 .
..
... ... ..
..
 .
.
.
.
 .


0 . . . . . . . . . 0 A(k)










.









Notice from the definition of vec[M ] and the block diagonal structure of (In ⊗ P ),
that the j-th block in (In ⊗ P ) corresponds to mean first passage times, mi1 i2 ,j ,
to node j. Since a permutation matrix simply acts as a relabeling of elements,
assume without a loss of generality, that the elements associated with B vary from
74
Chapter 3. The Hitting Time of Multiple Random Walks
i1 , i2 ∈ {1, . . . , r}. Therefore, the equations associated with B have the form
mi1 i2 ,j = 1+
X
(1)
(2)
mk1 k2 ,j pi1 ,k1 pi2 ,k2 −
k1 ,k2
−
X
X
(1)
(2)
mjk2 ,j pi1 ,j pi2 ,k2
k1
k2 6=j
(1)
(2)
(1)
(2)
mk1 j,j pi1 ,k1 pi2 ,j − mjj,j pi1 ,j pi2 ,j
k2
k1 6=j
= 1+
X
(1)
(2)
mk1 k2 ,j pi1 ,k1 pi2 ,k2 ,
k1 ,k2 6=j
where the subtracted terms in the expression above are associated with entries of
E (i.e., to the columns of P that are set to zero). If all subtracted terms are zero for
each mi1 i2 ,j , i1 , i2 ∈ {1, . . . , r}, this implies that there exists no i1 , i2 ∈ {1, . . . , r}
(1)
(2)
such that pi1 j > 0 or pi2 j > 0; that can only be true if there exists no path to node
j from any node i1 or i2 , which is impossible by definition of each P (i) . Therefore,
for each irreducible row-stochastic component of P¯ , there is at least one non-zero
element E such that a column of that component is set to zero, allowing us to
apply Lemma 18.
Now, proceeding to the N -agent case, similar to the 2 agent case, we require
that a unique solution exists for the N agent case if (InN +1 − (In ⊗ P )(InN +1 − E))
from (3.12) is invertible. For this to hold true it must be that (In ⊗ P )(InN +1 − E)
has |λmax | < 1. Similar to before, given that In ⊗ P and hence (In ⊗ P )(InN +1 −E)
is a block diagonal matrix, we need only show that each block has |λmax | < 1.
The proof continues a parallel line of argument to the 2-agent case shown in
75
Chapter 3. The Hitting Time of Multiple Random Walks
Theorem 23. More explicitly, each P can be deconstructed into block upper
triangular matrix composed of square matrices along the diagonal, each of which
is irreducible. Since P is row-stochastic, then each irreducible block is either
row-stochastic or sub-stochastic. If the irreducible block is row-stochastic then
a column of that block is necessarily set to zero due to the connectivity of the
graph, and hence by Lemma 18 has |λmax | < 1. If the irreducible block is not
row-stochastic then it is necessarily sub-stochastic, and hence |λmax | < 1.
Given this representation of the N -agent group hitting time, a natural question
is whether one can determine a simplified expression for this quantity which is a
function of the eigenvalues of P (h) , similar to the expression in Theorem 1. As
mentioned earlier, proof of that theorem relies on the ability to extort knowledge
of Md . Consider for example the 2-agent case; if we try to find the entries of Md in
a similar fashion to the single agent case by pre-multiplying (3.9) with π (1) ⊗ π (2)
we have that
1Tn = (π (1) ⊗ π (2) )T Md .
This is a system of n equations and 2n − 1 unknowns, and thus the solution of Md
is under determined. Therefore, even though one may be able to express the group
hitting time as a function of the eigenvalues, it is currently not well understood
how.
76
Chapter 3. The Hitting Time of Multiple Random Walks
3.3.2
Random-walkers covering subgraphs
The group hitting time as stated thus far is of interest in its own right. However, in many applications it is often desirable to have a notion of the same quantity when multiple agents don’t have access to the entire graph. In the following
section we tackle this problem by utilizing reducible graphs. We will leverage the
framework of the N -agent group hitting time for irreducible Markov chains in
order to generalize our results to reducible Markov chains.
Consider h ∈ {1, . . . , N } undirected irreducible weighted graphs G (h) = (V (h) , E (h) , P (h) )
(h)
with node sets V (h) ⊂ {1, . . . , n} such that ∪N
= {1, . . . , n}. The edge sets
h=1 V
(h)
satisfy E (h) ⊂ V (h) × V (h) and have corresponding weight matrices P (h) = [pi,j ] for
(h)
(h)
h ∈ {1, . . . , N } with the property pi,j ≥ 0 if (i, j) ∈ E (h) and pi,j = 0 otherwise.
As before, each matrix P (h) describes a Markov chain on the graph.
(h)
Let Xt
∈ V (h) denote the location of N random walkers at time t ∈ {0, 1, 2, . . . }.
For any N + 1 nodes ih ∈ V (h) and j ∈ {1, . . . , n}, the first passage time from any
ih , h ∈ {1, . . . , N } to j, denoted by Ti1 ...iN ,j , is the first time that any random
walker reaches node j, when starting from nodes ih , h ∈ {1, . . . , N } and is given
by
(h)
Ti1 ...iN ,j = min{t ≥ 1 | For h ∈ {1, . . . , N }, any Xt
77
(h)
= j given that X0
= ih }.
Chapter 3. The Hitting Time of Multiple Random Walks
Similar to case where each agent has access to the entire graph, we have the
following Lemma, whose proof is equivalent to that of Lemma 22.
Lemma 24. (Recursive formulation of first passage time for multiple randomwalkers over subgraphs): Consider the graphs G (h) = (V (h) , E (h) , P (h) ) satisfying
(h)
= {1, . . . , n} and let P (h) be the transition matrix associated
the property ∪N
h=1 V
with G (h) . Also, let |V (h) | denote the cardinality of each node set and let mi1 i...iN ,j =
E[Ti1 ...iN ,j ] denote the first passage time from any ih ∈ V (h) to j ∈ {1, . . . , n}, then
mi1 ...iN ,j = 1 +
X
(1)
(N )
mk1 ...kN ,j pi1 ,k1 . . . piN ,kN ,
k1 ,...,kN 6=j
or, in matrix notation,
M = 1α 1Tn + (P (1) ⊗ · · · ⊗ P (N ) )M − (P (1) ⊗ · · · ⊗ P (N ) )Md ,
where α =
QN
h=1
(3.14)
|V (h) |, M = [mi1 ...iN ,j ] and Md = [δi1 ...iN ,j mi1 ...iN ,j ].
Proof. The formulation of this system follows in a similar fashion the N agent
case, with the exception that now, if node ih has the property that ih ∈
/ V (h) ⊂
{1, . . . , n} then the corresponding mi1 ...iN ,j value is zero.
Given the formulation for the N -agent first passage time matrix over multiple
subgraphs, we now determine the average first visit time to any node in the full
graph. Like before, first we calculate the relative frequency of being at any given
node. Unlike before, however, each agent operates over a subset of the nodes in
78
Chapter 3. The Hitting Time of Multiple Random Walks
the graph. Let π (h) be the stationary distribution associated with Markov chain
P (h) ∈ Rr×r , where r ≤ n, and for convenience assume that V (h) ⊂ {1, . . . , n}
˜ (h) be the vector
is an ordered ascending set (i.e., V h = {3, 9, 20}). Then, let π
˜ (h)(h) = π (h)
˜ (h)
whose entries are given by π
for i ∈ {1, . . . , |V (h) |} and π
= 0
i
i
Vi
˜ (h) corresponds to the stationary distribution of each
otherwise. In other words, π
agent over the entire graph, not just its subgraph. Therefore if an agent never
˜ (h) ,
visits a node, its visit frequency to that node is 0. Given the padded vector π
˜ ave =
we write the average visit frequency as π
PN
˜
h=1 (π
(h)
)/N . Notice that this
interpretation of average visit frequency takes into account that multiple Markov
chains are running in parallel.
Now, the N -agent mean first passage time from start nodes ih , h ∈ {1, . . . , N },
denoted hi1 ...iN , is given by
hi1 ...iN =
n
X
˜ ave,j .
mi1 ...iN ,j π
j=1
Therefore, the average time to go from any set of N nodes to a single node in a
graph is given by
|V (1) |
HN =
X
i1 =1
|V (N ) |
(1)
π i1
···
X
(N )
π iN
iN =1
n
X
˜ ave,j
mi1 ...iN ,j π
j=1
˜ ave
= (π (1) ⊗ · · · ⊗ π (N ) )M π
˜ ave ⊗ π (1) ⊗ · · · ⊗ π (N ) ) vec[M ],
= (π
where as before, we denote HN as the N -agent group hitting time.
79
(3.15)
Chapter 3. The Hitting Time of Multiple Random Walks
We are now ready to state our main result.
Theorem 25. (Group hitting time for irreducible subgraphs): Consider the N
(h)
graphs G (h) = (V (h) , E (h) , P (h) ) satisfying the property ∪N
= {1, . . . , n} and
h=1 V
let P (h) ∈ R|V
(h) |×|V (h) |
be the irreducible transition matrices associated with G (h) .
P
˜ ave = ( N
˜ (i) )/N , and let E ∈ Rαn×αn be the diagonal matrix which
Also, let π
i=1 π
satisfies the equality E vec[M ] = vec[Md ]. Then the following hold:
(i) the group hitting time of the Markov chain is given by
˜ ave ⊗ π (1) ⊗ · · · ⊗ π (N ) )T vec[M ],
HN = (π
where
(3.16)
vec[M ] = (Iαn − (In ⊗ P (1) ⊗ · · · ⊗ P (N ) )(Iαn − E))−1 1αn ,
and α =
QN
h=1
|V (h) |, and
(ii) the hitting time between nodes h1 , . . . , hN and k, denoted mh1 ...hN ,k , of the
Markov chain is given by
...hN ,k T
mh1 ...hN ,k = vec[[Iih11...i
]] vec[M ].
N ,j
Proof. The proof of this theorem follows the exact same logic as the proof of
Theorem 23
This theorem immediately leads to the following corollary.
Corollary 26. (Group hitting time for reducible Markov chains): Consider N
multiple Markov chains, each with a transition matrix P (h) ∈ Rn×n satisfying
80
Chapter 3. The Hitting Time of Multiple Random Walks
the property that if P (h) is reducible, then there exists a permutation matrix S (i)
such that (S (h) )T P (h) S (h) is block diagonal with exactly one irreducible component.
P
(h)
Let π ave = ( N
)/N , with the property that π ave,j 6= 0 for all j. Also, let
h=1 π
E ∈ Rn
N +1 ×nN +1
be the diagonal matrix which satisfies the equality E vec[M ] =
vec[Md ]. Then the following properties hold:
(i) the group hitting time of the Markov chain is given by
HN = (π ave ⊗ π (1) ⊗ · · · ⊗ π (N ) )T vec[M ],
where
vec[M ] = (InN +1 − (In ⊗ P (1) ⊗ · · · ⊗ P (N ) )(InN +1 − E))−1 1nN +1 , and
(ii) the hitting time between nodes h1 , . . . , hN and k, denoted mh1 ...hN ,k , of the
Markov chain is given by
...hN ,k T
mh1 ...hN ,k = vec[[Iih11...i
]] vec[M ].
N ,j
Proof. First, note that the Kronecker product of two block diagonal matrices
generates a block diagonal matrix. Second, notice from Definition 17 property (i)
that
(S (1) )T P (i) S (1) ⊗ · · · ⊗(S (N ) )T P (i) S (N ) =
((S (1) )T ⊗ · · · ⊗(S (N ) )T )(P (1) ⊗ · · · ⊗ P (N ) )(S (1) ⊗ · · · ⊗ S (N ) ),
and so there exists a permutation matrix (S (1) ⊗ · · · ⊗ S (N ) ) that makes (P (1) ⊗ · · · ⊗ P (N ) )
block diagonal. Since exactly one block from each matrix is not exactly zero, the
81
Chapter 3. The Hitting Time of Multiple Random Walks
same is true of (S (1) )T P (i) S (1) ⊗ · · · ⊗(S (N ) )T P (i) S (N ) . This block corresponds to
matrix P in Theorem 25. The rest of the proof follows by noticing that each node
in the graph is reached if and only if π ave , j 6= 0 for all j. This is due to the
(h)
fact that π k 6= 0 when k denotes a persistent reachable node in the graph, and
(h)
π k = 0 otherwise.
3.3.3
Computational Complexity
Due to the extensive use of Kronecker products, it is important to verify the
memory and computational costs of the group hitting time. Looking at equation (3.11), we can assert that the group hitting time is affected by the curse
of dimensionality; with n nodes and N agents, the matrix In ⊗ P (1) ⊗ · · · ⊗ P (N )
contains n2N +2 elements. For example, given n = 100 nodes and N = 10 agents
the size of that matrix is 1044 × 1044 . The most intense operation is the inversion of In ⊗ P (1) ⊗ · · · ⊗ P (N ) , which requires a cost of O(k 3 ) where k is the
number of elements in the matrix [15], thus in our case this becomes O(n6N +6 ).
Noticing that the first Kronecker product in (3.12) is between the identity matrix
and the P = P (1) ⊗ · · · ⊗ P (N ) matrix, this implies In ⊗ P is block diagonal and
therefore we can store and invert single blocks, reducing memory cost to O(n2N )
and inversion cost to O(n6N ). For the more general hitting time formula described by equation (3.15), the complexity can be further reduced. Given (3.15)
82
Chapter 3. The Hitting Time of Multiple Random Walks
describes random-walks on subgraphs, then |V (h) | = nβh for some βh ∈ (0, 1],
and therefore the number of elements in the matrix Iαn ⊗ P (1) ⊗ · · · ⊗ P (N ) is
Q
2N +2
2
. This leads to a computational complexity for the inversion equal
( N
h=1 βh )n
Q
6
6N +6
to O ( N
β
)n
. Similar to before, we can take advantage of the fact
h=1 h
that the first Kronecker product in Iαn ⊗ P (1) ⊗ · · · ⊗ P (N ) is the identity matrix,
Q
Q
N
6
6N
2
2N
β
)n
,
β
)n
and
O
(
reducing memory and inversion costs to ( N
h=1 h
h=1 h
respectively.
In special circumstances, the above computational complexity can be dramatically reduced. This happens when the intersection between a subgraph of a single
agent, does not intersect with any other agents’ subgraph. In this case, the disjoint agent’s hitting time over its subgraph can be calculated independently of the
group hitting time of the other N − 1 agents. The single agent hitting time can
then be averaged with the N − 1 agent group hitting time to generate the N agent
group hitting time. In the case where every agent owns its own disjoint region,
the computational complexity scales to O(
PN
h=1
|V (h) |12 ), or to O(
PN
h=1
|V (h) |6 )
when exploiting the Kronecker product between the identity matrix and a transition matrix as was done previously. It is clear that uniformly partitioning the
graph between agents, or partitioning as close to uniform as is possible, we get
the lowest computational complexity.
83
Chapter 3. The Hitting Time of Multiple Random Walks
In the next section we compute an optimized group hitting time for various
graph topologies.
3.4
Numerical optimization of the group hitting
time
In the following sections we study the transition matrices that arise from the
numerical optimization of the group hitting time. In particular, we look at the
minimization of (3.16) described by Problem 6 below. The problem is numerically
solved using a sequential quadratic programming solver as implemented by MATLAB’s fmincon optimization algorithm, details of which are discussed in section
3.4.3.
(i)
Problem 6. (Group hitting time minimization): Let H1 denote the single agent
hitting time for random-walker i ∈ {1, . . . , N }. Given the stationary distributions
π 1 , π 2 , . . . , π N and Graph G with vertex set V and edge set E, find the transitions
84
Chapter 3. The Hitting Time of Multiple Random Walks
matrices P (1) , P (2) , . . . , P (N ) solving:
minimize
˜ ave ⊗ π)T (Iαn − (In ⊗(P (1) ⊗ P (2) · · · ⊗ P (n) ))(Iαn − E))−1 1αn
(π
subject to P (i) 1|V (i) | = 1|V (i) | , for each i ∈ {1, . . . , N }
(π (i) )T P (i) = (π (i) )T , for each i ∈ {1, . . . , N }
(i)
0 ≤ ph,k ≤ 1, for each (h, k) ∈ E and i ∈ {1, . . . , N }
(i)
ph,k = 0, for each (h, k) ∈
/ E and i ∈ {1, . . . , N }
P (i) is irreducible for i ∈ {1, . . . , N }.
The constraints in Problem 6, including the final one on the irreducibility of
P (i) , guarantee that the conditions of Theorem 25 are satisfied. In practice, it is
hard to enforce the irreducibility constraint during each step of an iterative optimization algorithm; our approach is to relax the constraint and verify a posteriori
that the iteratively-computed solution satisfies the irreducibility constraint. In all
the computational settings we considered, we never encountered a solution that
violated the irreducibility constraint.
In order to build intuition on the Markov chain combinations that generate
optimal group hitting time values, we present numerical results for the ring graphs,
complete graph and lattice graph shown in Figure 3.1. We look at two cases in
particular; one in which every random walker is required to visit all nodes in
the graph, and one in which random walkers are allowed to visit subgraphs. For
˜ ave to be uniform.
simplicity, we always restrict the stationary distribution π
85
Chapter 3. The Hitting Time of Multiple Random Walks
Figure 3.1: From left to right, example of a 5 node ring, 5 node complete, 9
node lattice and 4 node ring graph with self-loops.
3.4.1
Random-walkers covering the full graph
In the following we study which Markov chains generate optimal group hitting
time values when every random walker must visit every node in the graph. Surprisingly in fact, we will observe the random walks that generate optimal group
hitting times can be different. In each example we define individual agent’s stationary distribution as the uniform distribution. It is easily verified that with this
˜ ave,j = π
˜ ave,k for all j, k is always
choice of stationary distribution, the condition π
met no matter how many agents are added to the system.
We begin with the ring graph. For a singe random walker, the transition
matrix which generates the minimal hitting time is simply the one describing a
cycle (i.e, moving to a neighboring node with probability 1), and for the 5 node
ring graph shown in Figure 3.1. This is as expected as a cycle describes the fastest
time to reach any node from any other node. It turns out, for the multi-agent case
the optimal group hitting time occurs when every agent performs its own cycle; an
86
Chapter 3. The Hitting Time of Multiple Random Walks
example of this for three random-walkers is shown in Figure 3.2. Moreover, since
the group hitting time averages over all potential initial conditions, the direction
of cycles does not matter. In other words, one agent can go clockwise while the
other goes counter-clockwise. A summary of the optimal group and individual
random walker hitting times on a ring graph shown in Figure 3.1 is given in Table
3.1.
Figure 3.2: Probability to move along each edge of a ring graph (left) and
complete graph (right) for 3 random walkers. In the case above, the probability
to move along each edge is 1 which indicates a cycle. The group hitting time for
the shown trajectories for both ring and complete graph are H3 = 1.8
Random Walker(s)
One
Two
Three
Red
3.0
3.0
3.0
Blue
–
3.0
3.0
Green HN
–
3
–
2.2
3.0
1.8
Table 3.1: Group hitting time values for random walks shown in Figure 3.3. The
last column indicates the group hitting time for each case whereas the middle
three columns indicate each random-walkers individual hitting time. Surprisingly,
the ring and complete graph exhibit equivalent hitting time results.
Given the results for a ring graph, one can imply what will happen for the
complete graph for a single agent. As it happens, since a cycle exists, this is also
87
Chapter 3. The Hitting Time of Multiple Random Walks
the optimal strategy for the complete graph for both the single and multi-random
walker cases. Again direction of cycle does not matter as is seen for the three
agent case shown in Figure 3.2. A summary of the optimal group and individual
random walker hitting times on a complete graph shown in Figure 3.1 is equivalent
to the ring graph and thus is given in Table 3.1.
Next, we look at the lattice graph. Figure 3.3 shows the optimal trajectories
found ranging from a single random walker up to 3 random walkers. It is interesting to note that individual agents trajectories are quite different unlike for the
ring and complete graphs. This can be more easily seen by observing each agent’s
individual hitting time as shown in Table 3.2. What is surprising is that the transition matrices that generate the optimal group hitting time for the multi-random
waker cases are in fact sub-optimal individually. Interestingly enough, if one substitutes the optimal single agent transition matrix from the single random walker
case in for any/all of the multi-waker transition matrices, the group hitting time
becomes worse for those multi-waker cases.
Random Walker(s)
One
Two
Three
Red
6.8
7.7
7.0
Blue
–
10.5
15.9
Green HN
–
6.8
–
4.1
16.9
2.9
Table 3.2: Group hitting time values for random walks shown in Figure 3.3. The
last column indicates the group hitting time for each case whereas the middle
three columns indicate each random-walkers individual hitting time.
88
Chapter 3. The Hitting Time of Multiple Random Walks
Figure 3.3: Probability to move along each edge of a lattice graph for 1 random
walker (left), 2 random walkers (middle) and 3 random walkers (right). In each
graph, the opacity of a line indicates the probability to move along an edge.
Thus far we have seen that repeating multiple copies of the fastest random
walk is not always the most optimal. In fact, through simulation we’ve seen that
repeating random walks is only optimal when a cycle is the optimal single agent
strategy. In the following section we explore in more detail how the optimal group
hitting time is affected when working with sub graphs.
3.4.2
Random-walkers covering subgraphs
In the following section we build intuition on how the group hitting time is
affected over subgraphs. We will observe that working with subgraphs sometimes
improves the group hitting time, but can also cause the group hitting time to
worsen. We look at two cases in particular, when the subgraphs overlap and
when they do not (i.e., the nodes are partitioned amongst random walkers). The
subgraphs are not necessarily allocated in any optimal way, they are simply chosen
so that the stationary distribution is uniform over all nodes. For the simple
89
Chapter 3. The Hitting Time of Multiple Random Walks
examples shown, the stationary distributions for each random walker are defined
(i)
˜ ave,j /Nj where Nj denotes the number of agents who share node j. For
as π j = π
(i)
˜ ave,j if node j is only owned by one agent. For comparison with
example, π j = π
results presented in the previous section, we work with the ring and lattice graphs
shown in Figure 3.1.
The case when subgraphs overlap is studied for the 5 node ring and 9 node
lattice graph, the results of which can be seen in Figure 3.4. For the ring graph
shown, the optimal group hitting time is H2 = 2.5 in contrast to H2 = 2.2 from full
graph case (Table 3.1). For the lattice graph, the group hitting time is H2 = 3.6
in contrast to H2 = 4.1 (Table 3.2). Therefore, from these two examples, we see
that each agent covering less nodes is not always indicative of lower group hitting
time.
Figure 3.4: Probability to move along each edge of a 5 node ring graph with two
agents (left), and 9 node lattice graph with two agents (right). In each graph, the
opacity of a line indicates the probability to move along an edge.
The case when subgraphs are partitioned is analyzed for the 9 node lattice and
4 node ring graph shown in Figure 3.5. For the ring graph we see that the group
90
Chapter 3. The Hitting Time of Multiple Random Walks
hitting time is H2 = 1.5 whereas for the case where each agent covers the whole
graph H2 = 1.9; we do not show the figure for the latter case, however, recall that
the optimal full graph trajectory is simply a cyclic tour for each agent. Now, for
the partitioned lattice graph, the group hitting time is H3 = 3.7 in contrast to
H3 = 2.9 (Table 3.2). As before, we see that each agent covering less nodes, which
are partitioned, is not indicative or lower group hitting time.
Figure 3.5: Probability to move along each edge of a 4 node ring graph with two
agents (left), and 9 node lattice graph with three agents (right). In each graph,
the opacity of a line indicates the probability to move along an edge.
Clearly, the small sample study presented here leaves open many future avenues
that can be explored when attempting to optimize the group hitting time. For
example, notice that one can not always partition a graph and achieve an arbitrary
˜ ave . Also, it’s unclear what happens when you allow π
˜ ave and therefore individual
π
stationary distributions to vary. We leave these, among other questions to future
work.
91
Chapter 3. The Hitting Time of Multiple Random Walks
3.4.3
Implementation Notes
In order to generate repeatable and accurate results, we utilize a sequential
quadratic programming (SQP) solver as implemented in MATLAB’s fmincon optimization algorithm. Several other solvers are available, however, the SQP solver
has the most desirable mathematical properties. More specifically, with the SQP
solver, given that the maximum number of iterations is not reached, the minimization algorithm stops when the first-order necessary Karush-Kuhn-Tucker conditions are approximately satisfied; conditions are satisfied in the norm-sense with a
tolerance of 10−6 . For all results used, the first-order optimality stopping criteria
were satisfied.
The group hitting time results presented were taken as the minimum of 1000
optimization runs, each starting from a random initial conditions. Surprisingly,
with the exception of the lattice, every initial condition converged to the minimum
group hitting time value (within a 10−6 to 10−11 tolerance). For the two and three
agent results presented in Table 3.2, we found that solutions converged to the
minimum group hitting time value within a 10−2 tolerance for 96 and 93 percent
of samples, respectively. Therefore, we claim with reasonable confidence that a
local minimum for each group hitting time value was reached, if not a global
minimum. On a desktop computer with an Intel i7-4790 processor and 8 Gb of
RAM running MATLAB 2014b, the lattice with 9 nodes and 3 agents shown in
92
Chapter 3. The Hitting Time of Multiple Random Walks
Figure 3.2 took the longest time to run, averaging 10 minutes per run, whereas all
other simulations took less than a minute to fractions of a second per case run.
3.5
Summary
We have studied the hitting time of multiple random walkers on a graph and
have presented the first formulation of this quantity, which we denote the group
hitting time. Moreover, we have presented the first closed form solution for calculating the first hitting time between any specified set of nodes in a graph for
both the single and multi-agent cases. Finally, we posed the group hitting time
as an optimization problem and provided detailed simulation results which help
to build insight into the transition matrices that minimize this quantity.
This concludes our discussion of the hitting time. In the next chapter we begin
our discussion on multi-agent coverage control.
93
Chapter 4
Partitioning with
One-to-Base-Station
Communication
In this chapter and the following we discuss problems related to partitioning
and coverage control. The chapter is organized as follows. In Section 4.1 we setup
preliminary notation, introduce the concept of partitions and coverings, introduce
the one-to-base station network model and present our problem in technical detail.
In Section 4.2 we introduce some novel cost functions and present the solution
to our problem. In Section 4.3 we discuss implementation issues and present
simulation results. In the final section we summarize our findings.
94
Chapter 4. Partitioning with One-to-Base-Station Communication
4.1
Preliminaries and problem statement
The one-to-base-station communication model studied in this paper requires
that, in the design of coverage algorithms, we adopt overlapping coverings instead
of partitions. In this Section we translate concepts used in partitioning of continuous environments [13] to coverings on graphs. In our notation, R>0 , R≥0 , and
Z≥0 respectively denote the sets of positive, nonnegative and non-negative integer
numbers. Given a set A, |A| denotes the number of elements in A. Given sets
A, B, their union and intersections are denoted as A ∪ B and A ∩ B, respectively,
and their difference is A \ B = {a ∈ A | a ∈
/ B}.
4.1.1
Graphs and Distances
Let the finite set Q be a set of points in a continuous environment. These
points can represent locations or small areas of interest. They are assumed to
be the nodes of an (undirected) weighted graph G(Q) = (Q, E, Ω) with edge set
E ⊂ Q × Q and, with a slight abuse of notation, weight matrix Ω = [Ωi,j ] with the
property that Ωi,j = Ωj,i > 0 if (i, j) ∈ E and Ωi,j = 0 otherwise. We assume that
G(Q) is connected and think of the Ωi,j edge weights as travel distances between
nearby nodes.
95
Chapter 4. Partitioning with One-to-Base-Station Communication
In any weighted graph G(Q) there is a standard notion of distance between
vertices defined as follows. A path in G is an ordered sequence of vertices such
that any consecutive pair of vertices is an edge of G (i.e., if (i, j) ∈ E). The weight
of a path is the sum of the weights of the edges in the path. Given vertices h
and k in G, the distance between h and k, denoted dG (h, k), is the weight of the
lowest weight path between them, or +∞ if there is no path. If G is connected,
then the distance between any two vertices is finite. By convention, dG (h, k) = 0
if h = k. Note that by definition of weights Ωi,j that dG (h, k) = dG (k, h), for any
h, k ∈ Q.
4.1.2
Coverings of Graphs
We will be covering Q with n subsets or regions which will each be owned by
an individual agent.
Definition 27 (n-Covering). Given the graph G(Q) = (Q, E, Ω), we define a
n-covering of Q as a collection P = {Pi }ni=1 of subsets of Q such that:
(i)
Sn
i=1
Pi = Q;
(ii) Pi 6= ∅ for all i ∈ {1, . . . , n};
Let Covn (Q) to be the set of n-coverings of Q.
96
Chapter 4. Partitioning with One-to-Base-Station Communication
Note that a vertex in Q may belong to multiple subsets in P , i.e., a vertex
may be covered by multiple agents. The above definition is an important change
from prior work [26], which was limited to partitions of Q, defined as follows.
Definition 28 (n-Partition). A n-partition is a n-covering with the additional
property that:
(iii) if i 6= j, then Pi ∩ Pj = ∅.
Let Partn (Q) to be the set of n-partitions of Q.
Among the ways of covering Q, there is one which is worth special attention. Before we state the partition, let us define the vector of weights w :=
{w1 , . . . , wn }, such that wi > 0 and
W = {w ∈ Rn>0 |
Pn
i=1
Pn
j=1
wi = 1.
For brevity, we denote
wi = 1}. Then given w ∈ W and a vector of dis-
tinct points c ∈ Qn , the partition P ∈ Partn (Q) is said to be a multiplicatively
weighted Voronoi partition of Q generated by c and weighted by w if, for each
Pi and all k ∈ Pi , we have ci ∈ Pi and
1
d (k, ci )
wi G
≤
1
d (k, cj ),
wj G
for j 6= i. The
elements of c are said to be the generators of the Voronoi partition multiplicatively
weighted by w. Note that there can be more than one multiplicatively-weighted
Voronoi partition generated by c and w since how to assign tied vertices is unspecified. Multiplicatively-weighted Voronoi partition allow us to accommodate
heterogeneous agents. For example, if agent i is faster than another agent j (i.e.,
97
Chapter 4. Partitioning with One-to-Base-Station Communication
wi > wj ), it would make sense that agent i should control more territory than
agent j. Multiplicatively-weighted Voronoi partitions are a subset of generalized
Voronoi partitions, which will be discussed in further detail in Chapter 5 From
this point forward we refer to multiplicatively-weighted Voronoi partitions simply as Voronoi partitions and the vector of weights w is given and fixed. Given
that weights are fixed, for the rest of the paper we refer to a Voronoi partition
generated by c and w simply as a Voronoi partition generated by c.
Given the above, we are now ready to state the one-to-base station network
model.
4.1.3
One-to-Base-Station Robotic Network Model
Given a team of n robotic agents and a central base station, each agent i ∈
{1, . . . , n} is required to have the following basic computation capabilities:
(C1) agent i can identify itself to the base station; and
(C2) agent i has a processor with the ability to store a region Si ⊂ G(Q) and a
center si ∈ Si .
Each i ∈ {1, . . . , n} is assumed to communicate with the base station according to
the asynchronous one-to-base-station communication model described as follows:
98
Chapter 4. Partitioning with One-to-Base-Station Communication
(C3) there exists a finite upper bound ∆ on the time between communications
between i and the base station. For simplicity, we assume no two agents
communicate with the base station at the same time.
The base station must have the following capabilities:
(C4) it can store an arbitrary n-covering of Q, P = {Pi }ni=1 , a list of locations
c ∈ Qn and weights w ∈ W;
(C5) it can perform computations on subgraphs of G(Q); and
(C6) it can store and operate on multiple n-coverings of Q, P = {Pi }ni=1 and a list
of locations c ∈ Qn .
4.1.4
Problem Statement
We are now ready to state our problem of interest.
Given weights w ∈ W assume that, for all t ∈ R≥0 , each agent i ∈ {1, . . . , n}
maintains in memory a subset Si (t) of environment Q and a vertex si (t) ∈ Si (t).
Our goal is to iteratively update the covering S(t) = {Si }ni=1 and the centers
s(t) = {si }ni=1 while solving the optimization problem:
min
min
s∈Qn S∈Covn (Q)
U (s, S),
(4.1)
for some cost function U (s, S) subject to the constraint that every node in the
environment Q maintains coverage from some agent, and subject to the constraint
99
Chapter 4. Partitioning with One-to-Base-Station Communication
imposed by the robot network model with asynchronous one-to-base-station communication.
4.2
Proposed Solution
In this section we present our proposed solution to the problem described in
Section4.1.4. We begin by first introducing some useful cost functions and their
properties. Then, we present an algorithm that uses these cost functions and show
that it solves (4.1).
4.2.1
Cost Functions
Let weight function φ : Q → R>0 be a bounded positive function which assigns
a relative weight to each element of Q. The weight assigned to a node by φ can
be thought of as the ”service time” or importance of that node. The one-center
function H1 gives the cost for a robot to cover a subset A ⊂ Q from a vertex
h ∈ A with relative prioritization given by φ:
H1 (h; A) =
X
dG (h, k)φ(k).
k∈A
This cost function leads us to the following definition.
100
Chapter 4. Partitioning with One-to-Base-Station Communication
Figure 4.1: The left image shows a grid environment whose corresponding graph
representation is shown in the right image. Each cell in the grid represents a
node in the graph and if two cells are adjacent, then there is an unit-weight edge
between those nodes. The black nodes in the graph denote the set of generalized
centroids for the corresponding grid environment.
Definition 29 (Centroid). We define the set of generalized centroids of A ⊂ Q
as the set of vertices in A which minimize H1 , i.e.,
C(A) := argmin H1 (h; A).
h∈A
In what follows, we drop the word “generalized” for brevity. Note that the
centroid of a set always belongs to the set. Figure 4.1 shows an illustrative example
of the set C(A) for a simple environment.
With these notions, we are ready to define other useful cost functions. We can
define the multi-center function H : Qn × Covn (Q) → R≥0 to measure the cost for
n robots to cover a n-covering P from the vertices c ∈ Qn by
H(c, P ) =
n X
X
1
dG (ci , k)φ(k).
w
i
i=1 k∈P
i
Note that if wi = wj for all i, j, then the multi-center cost function above is
the same as in [26]. Furthermore, we define the minimum cost-to-cover mapping
101
Chapter 4. Partitioning with One-to-Base-Station Communication
Hmin : Qn × Covn (Q) → R≥0 by
Hmin (c, P ) =
X
k∈Q
min
i
1
dG (ci , k) | k ∈ Pi φ(k).
wi
Note that if P is a partition, then Hmin (c, P ) = H(c, P ) for any c. We aim to
minimize these performance functions with respect to both the covering P and
the vertices c. In the motivational scenario we are considering, each robot will
periodically be asked to perform a task somewhere in its region with tasks located
according to distribution φ. When idle, the robots would position themselves at
the vertices c. By minimizing the coverage cost, the robot team minimizes the
expected distance between a task and the furthest robot which can service the
task.
We are almost ready to introduce a notion of optimal partition, the centroidal
Voronoi partition. Our discussion begins with the following results, which are
direct consequences of the above definitions.
Proposition 30 (Properties of H). Let P ∈ Partn (Q) and c ∈ Qn then the
following properties hold:
(i) If P 0 is a Voronoi partition generated by c, then
H(c, P 0 ) ≤ H(c, P ).
(ii) Let I ⊂ {1, . . . , n}. If c0 ∈ Qn satisfies c0i ∈ C(Pi ) for i ∈ I and c0j = cj for
j 6∈ I, then
102
Chapter 4. Partitioning with One-to-Base-Station Communication
H(c0 , P ) ≤ H(c, P )
with a strict inequality if ci ∈
/ C(Pi ) for any i ∈ I.
Proposition 31 (Properties of Hmin ). Let P 0 ∈ Partn (Q) be a Voronoi partition
generated by c ∈ Qn , then the following properties hold:
(i) If P ∈ Covn (Q) is a covering such that Pi0 ⊆ Pi for all i, then
Hmin (c, P 0 ) = Hmin (c, P ).
(ii) If P ∈ Partn (Q) is a partition satisfying P i ∩ Pi0 3 ci for all i, then
Hmin (c, P 0 ) ≤ Hmin (c, P ).
(iii) Let I ⊂ {1, . . . , n}. If c0 ∈ Qn satisfies c0i ∈ C(Pi0 ) for i ∈ I and c0j = cj for
j 6∈ I, then
Hmin (c0 , P 0 ) ≤ Hmin (c, P 0 )
with a strict inequality if ci ∈
/ C(Pi0 ) for any i ∈ I.
Propositions 30 and 31 imply that if P ∈ Partn (Q) and (c, P ) minimizes H
(equivalently Hmin ), then ci ∈ C(Pi ) for all i and P must be a Voronoi partition
generated by c. This motivates the following definition.
103
Chapter 4. Partitioning with One-to-Base-Station Communication
Figure 4.2: The figure shows two environments with two agents. Each cell
denotes a node in a graph and if two cells are adjacent, then there is a unit-weight
edge between those nodes. The left image shows a Voronoi partition generated by
the two agents. Note that the blue agent is not at its region’s centroid. The right
image is instead a centroidal Voronoi partition.
Definition 32 (Centroidal Voronoi Partition). P ∈ Partn (Q) is a centroidal
Voronoi partition of Q if there exists a c ∈ Qn such that P is a Voronoi partition
generated by c and ci ∈ C(Pi ) for all i.
For a given environment Q, a pair made of a centroidal Voronoi partition and
the corresponding vector of centroids is locally optimal in the following sense:
the cost functions H and Hmin cannot be reduced by changing either P or c
independently. Figure 4.2 demonstrates the difference between a Voronoi and
centroidal Voronoi partition.
4.2.2
The One-to-Base Coverage Algorithm
Given the cost function defined by U (s, S) and the One-to-Base Network model
described by (C1)–(C6), we introduce the One-to-Base Coverage Algorithm to
solve the optimization problem (4.1).
One-to-Base Coverage Algorithm
104
Chapter 4. Partitioning with One-to-Base-Station Communication
The base station maintains in memory an n-covering P = {Pi }ni=1 , vector of
locations c = (ci )ni=1 and normalized weights w = (wi )ni=1 , while each robot i
maintains in memory a set Si and a vertex si . The base station maintains in
temporary memory n-coverings P = {P i }ni=1 and P = {P i }ni=1 , along with vectors
c = (ci )ni=1 and c = (ci )ni=1 for computational purposes. At t = 0, let P (0) ∈
Covn (Q), S(0) = P (0), and let all ci (0)’s be distinct. Assume that at time
t ∈ R>0 , robot i communicates with the base station. Let P + , c+ , Si+ , and s+
i
be the values after communication. Then the base station executes the following
actions:
1: update P := P , c := c, P := P , c := c,
2: computesets
1
1
Pi,+ := x ∈ Q wi dG (x, ci ) < min wj dG (x, cj ) x ∈ Pj , j 6= i
1
Pi,− := x ∈ Pi ∩ ∪i6=j Pj wi dG (x, ci ) ≥ min w1j dG (x, cj ) x ∈ Pj , j 6=
i
3:
4:
5:
6:
7:
8:
update P i := (Pi \Pi,− ) ∪ Pi,+
for k ∈ Pi \c do
computesets 1
1
Pi,+ := x ∈ Q wi dG (x, k) < min wj dG (x, cj ) x ∈ Pj , j 6= i
1
Pi,− := x ∈ Pi ∩ ∪i6=j Pj wi dG (x, k) ≥ min w1j dG (x, cj ) x ∈ Pj , j 6=
i
update P i := (Pi \Pi,− ) ∪ Pi,+
update ci := k
if U c, P < U c, P then
update P i := P i
update ci := ci
+
11: Pi := P i
12: c+
i := ci
+
13: tell agent i to set Si+ := Pi+ and s+
i = ci
9:
10:
105
Chapter 4. Partitioning with One-to-Base-Station Communication
Remark 33 (Constant cost). Given the constant cost function U (c, P ) = α for
α ∈ R, for a set of initial conditions (c, P ), the One-to-Base Coverage Algorithm
produces a Voronoi partition generated by c.
Remark 34 (Full coverage). Notice that the set Pi,+ adds points to an agents
environment from other agent’s territory that are closer to it. Also, notice that
Pi,− only removes points from agent i’s territory if another agent is covering that
territory. Defining the sets in this way ensures that every point in the environment
will always have coverage by some agent.
We have the following main result on the limit behavior of the algorithm.
Theorem 35 (Convergence of One-to-Base Coverage Algorithm (Hmin )). Consider a network consisting of n robots endowed with the computation capacities
(C1), (C2) and communication capacity (C3), and a base station with capacities
(C4), (C5) and (C6). Assume the network implements the One-to-Base Coverage
Algorithm with U (c, P ) = Hmin (c, P ). Then the resulting evolution
(s, S) : R≥0 → Qn × Covn (Q)
converges in finite time to a pair (s∗ , S ∗ ) composed of a centroidal Voronoi partition S ∗ generated by s∗ .
106
Chapter 4. Partitioning with One-to-Base-Station Communication
Pareto-Optimal Partitions
Using the algorithms described thus far, ties along partitions’ boundaries are
not handled in any optimal way and can often be improved. The major source
of sub-optimal boundary allocation is due to the discrete nature of how centroids
of a region are selected. Often times when an agent has more than one “center”
location, the overall partition can become better balanced if the agent takes an
alternate center value as its centroid. The following definition and proposition
make this notion more precise.
Definition 36 (Pareto-Optimal Partition). Given a vector of positions c = {c1 , . . . , cn },
the Voronoi partition P generated by c is Pareto-optimal if for all c = {c1 , . . . , ci , . . . , cn }
for i ∈ {1, . . . , n} such that ci 6= ci and ci ∈ Pi , the Voronoi partition P generated
by c satisfies H(c, P ) ≤ H(c, P ).
As an immediate consequence of the definition of a Pareto-optimal partition
and of Proposition 30 we can conclude that every Pareto-optimal partition is also
a centroidal Voronoi partition. However, the reverse implication does not hold, as
shown in the following example.
Example 1 (One Dimensional Pareto-optimal Partition). Consider the three environments in Figure 4.3, each with two agents denoted by the colored circles.
Assume that each cell denotes a node in the graph and that unit-weight edges con107
Chapter 4. Partitioning with One-to-Base-Station Communication
nect any adjacent cells. Assume φ is constant. If the environment is partitioned
according to leftmost image of Figure 4.3, then each agent is at the centroid of its
region, and the graph is a centroidal Voronoi partition whose multi-center function cost-to-cover is H = 4. This partition is clearly not well balanced, and unless
ties are broken in some non-trivial way, this is a valid (worst-case) partition that
the system can reach. If however, the blue agent moves to its other centroid, as
shown in the middle image, then the worst case partition must be a variant of the
partition shown in the rightmost image whose cost-to-cover is H = 3. There exists
no partition with smaller cost-to-cover (w.r.t. H) by moving any single agent, and
hence the partition in the rightmost image is Pareto-optimal.
Figure 4.3: The figure shows three environments with two agents. Each cell
denotes a node in a graph, and if two cells are adjacent then there is an unitweight edge between those nodes.
The above results give that Pareto-optimal partitions are a subset of centroidal
Voronoi partitions. We can define the cost function, U (c, P ), in the One-toBase Coverage Algorithm such that the algorithm converges to a Pareto-optimal
108
Chapter 4. Partitioning with One-to-Base-Station Communication
partition. We define the new cost function
Hinf (c) =
X
k∈Q
min
i
1
dG (ci , k) | k ∈ Q φ(k).
wi
(4.2)
Notice that this function is different from Hmin in that it looks for the absolute
minimum distance to a point, k. The function Hinf allows the case when k ∈
/ Pi ,
but
1
d (c , k)
wi G i
<
1
d (c , k),
wj G j
for all j 6= i. The Hinf function is linked to the
multicenter function in the following sense.
Proposition 37 (Properties of Hinf ). Given c ∈ Qn and w ∈ W, let P be a
Voronoi partition generated by c, then
H(c, P ) = Hinf (c).
Proof. Voronoi partitions are optimal in the sense that H(c, P ) = Hinf (c) by
definition of Hinf .
We are now ready to state the main result of this subsection. Given the Oneto-Base Coverage Algorithm with U (s, S) = Hinf (c) we have the following result.
Theorem 38 (Convergence of One-to-Base Coverage Algorithm (Hinf )). Consider a network consisting of n robots endowed with the computation capacities
(C1), (C2) and communication capacity (C3), and a base station with capacities
(C4), (C5) and (C6). Assume the network implements the One-to-Base Coverage
109
Chapter 4. Partitioning with One-to-Base-Station Communication
Algorithm with U (c, P ) = Hinf (c). Then the resulting evolution
(s, S) : R≥0 → Qn × Covn (Q)
converges in finite time to a pair (s∗ , S ∗ ) composed of a Pareto-optimal partition
S ∗ generated by s∗ .
Some remarks are in order. First, it is possible for the One-to-base Algorithm
with U = Hmin to converge to a Pareto-optimal partition, however, it is not
guaranteed as in the case with U = Hinf . Second, if a partition is not Paretooptimal then the cost to cover a region, in the context of the multi-center function,
can be further decreased by making it Pareto-optimal. This point is clarified
in Proposition 37, which relates the multi-center function to Hinf . Finally, we
emphasize that the difference in the cost-to-cover a region for a Pareto-optimal
partition versus a centroidal Voronoi partition decreases as the map defining a
region becomes less coarse. This is because the notion of a centroidal Voronoi
partition not being Pareto-optimal only exists when a region has more than one
centroid, a property of discrete spaces but not of continuous ones. Therefore, as
the grid approximating a region becomes less coarse, the more likely it is that a
centroidal Voronoi partition is also Pareto-optimal.
110
Chapter 4. Partitioning with One-to-Base-Station Communication
Combining Hmin and Hinf
Although the One-to-Base Coverage Algorithm with U = Hinf is guaranteed
to converge to a Pareto-optimal partition whereas the algorithm with U = Hmin is
not, the algorithm with U = Hmin is still of practical importance. Computing Hinf
requires that for each node in the graph, all agents compare their relative distances
to that node regardless of whether that node exists in the agent’s territory or not.
The Hmin function, however, only requires a comparison if the node belongs in the
agent’s territory. Given the computational capabilities of the base station being
used, one method may be preferred over the other. A user can take advantage
of both algorithm properties by running the algorithm with U = Hmin until it
converges to get an initial partition, and then run the algorithm with U = Hinf
to reach a Pareto-optimal solution, if the solution system has not already reached
it during the U = Hmin portion of the algorithm. With a slight abuse of notation
we will refer to this combined algorithm as the One-to-Base Coverage algorithm
with U = Hmin,inf .
4.2.3
Convergence Proofs
In this section we prove Theorems 35 and 38. Any mention to the One-toBase Coverage Algorithm in this section will refer to the algorithm presented
in Section 4.2.2. Their proof is based on the following convergence result for
111
Chapter 4. Partitioning with One-to-Base-Station Communication
set-valued algorithms on finite state spaces, which can be recovered as a direct
consequence of [12, Theorem 4.3].
Given a set X, a set-valued map T : X ⇒ X is a map which associates to an
element x ∈ X a subset Z ⊂ X. A set-valued map is non-empty if T (x) 6= ∅ for
all x ∈ X. Given a non-empty set-valued map T , an evolution of the dynamical
system associated to T is a sequence {xn }n∈Z≥0 ⊂ X with the property xn+1 ∈
T (xn ) for all n ∈ Z≥0 .
Lemma 39 (Convergence under persistent switches). Let (X, d) be a finite metric
space. Given a collection of maps T1 , . . . , Tm : X → X, define the set-valued map
T : X ⇒ X by T (x) = {T1 (x), . . . , Tm (x)} and let {xn }n∈Z≥0 be an evolution of
T . Assume that:
(i) there exists a function U : X → R such that U (x0 ) < U (x), for all x ∈ X
and x0 ∈ T (x) \ {x}; and
(ii) for all i ∈ {1, . . . , m}, there exists an increasing sequence of times {nk | k ∈
Z≥0 } such that xnk +1 = Ti (xnk ) and (nk+1 − nk ) is bounded.
Let Fi = {x ∈ X | Ti (x) = x} be the set of fixed points of Ti . Then, for all
x0 ∈ X there exist N ∈ N and x¯ ∈ (F1 ∩ · · · ∩ Fm ) such that xn = x¯ for all n ≥ N .
Note that the existence of a common fixed point for the collection of maps Ti
is guaranteed by this result.
112
Chapter 4. Partitioning with One-to-Base-Station Communication
We now apply Lemma 39 to the evolution of One-to-base Coverage Algorithm
with U (c, P ) = Hmin (c, P ) and U (c, P ) = Hinf (c), respectively. To do so, for each
function given by U (c, P ), we must describe the algorithm as a set-valued map
and find a corresponding Lyapunov function. The first step is possible because
the One-to-Base Coverage Algorithm is well-posed in the sense of the following
immediate result.
Proposition 40 (Well-posedness). Let P ∈ Covn (Q) and c ∈ Qn such that ci ∈ Pi
and ci 6= cj for all i and all j 6= i. Then, P + and c+ produced by the One-to-Base
Coverage Algorithm meet the same criteria.
Given this result, the One-to-Base Coverage Algorithm can be written as a set
valued map. For any i ∈ {1, . . . , n}, we define the map TU,i : Qn × Covn (Q) →
Qn × Covn (Q) by
TU,i (c, P ) = {c1 , . . . , c+
i , . . . , cN },
{P1 , . . . , Pi+ , . . . , Pn } ,
+
where c+
i and Pi are defined per the algorithm when i is the communicating
robot, and U is dependent on the cost function we are referring to (i.e., U = Hmin
or U = Hinf ). Then we can define the set-valued map TU : Qn × Covn (Q) 7→
Qn × Covn (Q) by
TU (c, P ) = {TU,1 (c, P ), . . . , TU,N (c, P )}.
113
Chapter 4. Partitioning with One-to-Base-Station Communication
Thus, the dynamical system defined by the application of the algorithm is described by {c+ , P + } ∈ TU (c, P ).
For our Lyapunov arguments we will need to define M (P ) as the set of vertices
which are owned by multiple agents. We now proceed by stating two useful
propositions, which allow us to conclude Theorem 35.
Proposition 41 (Decaying Hmin cost function). After each iteration of the oneto-base station algorithm if (c+ , P + ) 6= (c, P ) then one of the following holds:
(i) Hmin (c+ , P + ) < Hmin (c, P ); or
(ii) Hmin (c+ , P + ) = Hmin (c, P ), and
|M (P + )| < |M (P )|.
Proof. If c+ = c then Hmin (c+ , P + ) ≤ Hmin (c, P ). This is a direct consequence
of how the sets Pi,+ and Pi,− are defined. Points are added to Pi if and only if
they are strictly closer to ci than any other center cj and hence the cost of Hmin
must decrease by the addition of these points. Points are removed if and only if
they are strictly farther away or tied points and so Hmin must decrease or stay the
same. If c+ 6= c then by lines 8-10 of the algorithm Hmin (c+ , P + ) < Hmin (c, P ).
For the case Hmin (c+ , P + ) = Hmin (c, P ) then for every x ∈ Pj \Pi , for all j 6= i,
there exists no point that is strictly closer to the center ci than any other center
114
Chapter 4. Partitioning with One-to-Base-Station Communication
cj , j 6= i. Therefore, no points can be added to Pi+ , and so if P + 6= P it must be
the case that |M (P + )| < |M (P )|.
Proposition 42 (Convergence of THmin ). The evolution of the One-to-Base Coverage Algorithm (c(t), P (t)) generated by the map THmin converges in finite time
to the intersection of the equilibria of the maps THmin ,i , that is, to a pair (c, P )
where P is a centroidal Voronoi partition generated by c.
Proof. The proof proceeds with an application of Lemma 39 to (c(t), P (t)). The
algorithm is the mapping THmin : Qn ×Covn (Q) 7→ Qn ×Covn (Q) defined above and
is well-posed. We can form a Lyapunov function using Proposition 41 as follows.
Since the set Q is finite, there exists only a finite number of possible values for Hmin
and |M |. Let m be the magnitude of the smallest non-zero difference between
any two values of Hmin . Let αM be larger than twice the maximum possible value
of |M |. Define V : Qn × Covn (Q) → R≥0 by
V (c, P ) = Hmin (c, P ) +
m
|M (P )|.
αM
Thanks to this scaling of |M (P )|, Proposition 41 implies that if (c0 , P 0 ) ∈ THmin (c, P ),
then either V (c0 , P 0 ) < V (c, P ) or (c0 , P 0 ) = (c, P ). Thus, V (c, P ) fulfills assumption (i) in Lemma 39. Moreover, the communication model (C3) assures that
assumption (ii) in Lemma 39 is met. Now, applying Lemma 39, we are assured
that the dynamics converge to a fixed point (c∗ , P ∗ ). It remains to show that P ∗
115
Chapter 4. Partitioning with One-to-Base-Station Communication
is a centroidal Voronoi partition generated by c∗ . We do this by refining in three
sequential steps the properties that the fixed point must have: P ∗ is a partition,
(c∗ , P ∗ ) is a a Voronoi partition, and finally (c∗ , P ∗ ) is a centroidal Voronoi par∗
tition. First, if P ∗ is not a partition, then Pi,−
6= ∅; this establishes that P ∗ is a
∗
partition. Second, if the partition P ∗ is not Voronoi, then Pi,+
6= ∅; this establishes
that (c∗ , P ∗ ) is a Voronoi partition. Third, if P ∗ is a Voronoi partition generated
/ C(Pi∗ ) for any i, then from part (iii) of Proposition 31 there exists
by c∗ , but c∗i ∈
∗
a location c∗∗
i (at a centroid location) that improves the cost to cover Pi ; line 4
of the algorithm guarantees that this location is checked, and lines 8-10 ensure
the position is updated to that location or one with an even lower cost. It should
be noted that an update in location to c∗∗
i can simultaneously lead to an update
∗
in territory Pi∗ : however, given that P ∗ is a partition, only Pi,+
can contribute to
the territory, which further decreases the coverage cost by the definition of Hmin
∗
. Therefore, we have established that there exists an update reducing the
and Pi,+
cost function THmin if the fixed point is not a centroidal Voronoi partition. Then,
P ∗ must be a centroidal Voronoi partition generated by c∗ . (Note that the fixed
point could also potentially have other properties in addition to being a centroidal
Voronoi partition, however, establishing those properties are beyond the scope of
this proof.)
116
Chapter 4. Partitioning with One-to-Base-Station Communication
Since updates to agent i in base-station memory also occur on the physical
agent, we can conclude the convergence proof of Theorem 35.
Finally, we state two propositions which allow us to conclude Theorem 38.
Proposition 43 (Decaying Hinf cost function). After each iteration of the oneto-base station algorithm if (c+ , P + ) 6= (c, P ) using Hinf as the cost function then
one of the following holds:
(i) Hinf (c+ ) < Hinf (c); or
(ii) Hinf (c+ ) = Hinf (c), and
Hmin (c+ , P + ) < Hmin (c, P ); or
(iii) Hinf (c+ ) = Hinf (c), Hmin (c+ , P + ) = Hmin (c, P ) and |M (P + )| < |M (P )|
Proof. If c+ = c then Hmin (c+ , P + ) ≤ Hmin (c, P ) and Hinf (c+ ) = Hinf (c). This is
a direct consequence of how we define Hinf and the sets Pi,+ and Pi,− . Points are
added to Pi if and only if they are strictly closer to ci than any other center cj
and hence the cost of Hmin must decrease by the addition of these points. Points
are removed if and only if they are strictly farther away or tied points and so
Hmin must decrease or stay the same. If c+ 6= c then by the lines 8-10 of the
algorithm Hinf (c+ ) < Hinf (c). For the case Hmin (c+ , P + ) = Hmin (c, P ) then for
every x ∈ Pj \Pi , for all j 6= i, there exist no point that is strictly closer to the
117
Chapter 4. Partitioning with One-to-Base-Station Communication
center ci than any other center cj , j 6= i. Therefore, no points can be added to
Pi+ , and so if P + 6= P it must be the case that |M (P + )| < |M (P )|.
Proposition 44 (Convergence of THinf ). The evolution of the One-to-Base Coverage Algorithm (c(t), P (t)) generated by the map THinf converges in finite time to
the intersection of the equilibria of the maps THinf ,i , that is, to a pair (c, P ) where
P is a Pareto-optimal partition generated by c.
Proof. The proof follows the lines of the proof of Proposition 42, with the important modification of using a different Lyapunov function, defined as follows. Let
i and m be the magnitude of the smallest possible non-zero difference between
two values of Hinf and Hmin , respectively. Let αm and αM be larger than twice the
maximum possible value of Hmin and |M |, respectively. If we define the function
V as
V (c, P ) = Hinf (c) +
i m
i
Hmin (c, P ) +
|M (P )|
αm
αm αM
and we invoke Proposition 43 and Lemma 39, we conclude that the dynamics
converge to a fixed point (c∗ , P ∗ ).
It remains to show that P ∗ is a Pareto-optimal partition generated by c∗ . If
∗
P ∗ is not a partition, then Pi,−
6= 0 and if the partition is not Voronoi then
∗
Pi,+
6= 0. Continuing by contradiction, assume that c∗ forms a Voronoi partition
which is not Pareto-optimal. This implies that there exists a c0 ∈ P ∗ where for at
118
Chapter 4. Partitioning with One-to-Base-Station Communication
least one agent c0i 6= c∗i and a partition P 0 generated by c0 such that H(c0 , P 0 ) <
H(c∗ , P ∗ ). By the definition of the algorithm and Proposition 37 this is not
possible. Therefore the fixed point partition is Pareto-optimal.
4.3
Implementation and Simulations
In order to efficiently implement the One-to-Base Coverage Algorithm and
under the assumption of using Hmin or Hinf as cost functions, we provide the
following revised version, which can be easily seen to be equivalent to that in
Section 4.2.2.
One-to-Base Coverage Algorithm – revised
The base station maintains in memory an n-covering P = {Pi }ni=1 , vector of
locations c = (ci )ni=1 and normalized weights w = (wi )ni=1 , while each robot
i maintains in memory a set Si and a vertex si . The base station maintains
in temporary memory an n-covering P = {P i }ni=1 and vectors c = (ci )ni=1 and
c = (ci )ni=1 for computational purposes. At t = 0, let P (0) ∈ Covn (Q), S(0) =
P (0), and let all ci (0)’s be distinct. Assume that at time t ∈ R>0 , robot i
communicates with the base station. Let P + , c+ , Si+ , and s+
i be the values after communication. Then the base station executes the following actions:
1:
2:
3:
4:
5:
6:
update P := P , c := c, c := c, P i = Q
for k ∈ Pi \c do
update ci := k
if U c, P < U c, P then
update ci := k
computesets 1
1
Pi,+ := x ∈ Q wi dG (x, ci ) < min wj dG (x, cj ) x ∈ Pj , j 6= i
119
Chapter 4. Partitioning with One-to-Base-Station Communication
Pi,−
i
7:
8:
9:
:= x ∈ Pi ∩ ∪i6=j Pj
1
d (x, ci )
wi G
≥ min
1
d (x, cj )
wj G
x ∈ Pj , j 6=
Pi+ := (Pi \Pi,− ) ∪ Pi,+
c+
i := ci
+
tell agent i to set Si+ := Pi+ and s+
i = ci
This revised version of the algorithm takes advantage of how the cost functions
Hmin and Hinf , and sets Pi,+ and Pi,− are defined. Indeed, setting P i = Q in line
1 avoids having to calculate Pi,+ and Pi,− for every k ∈ Pi \ci , as was done in
Section 4.2.2, because Hmin and Hinf already distribute costs to nodes that are
closer to one agent as opposed to another. Hence this implementation produces
the same evolutions but requires less memory, as we no longer need the set P , and
less computation time, as the sets Pi,+ and Pi,− are calculated only once.
We are now ready to proceed with our simulation results, which are obtained
by running the revised version of the algorithm. To demonstrate the utility of the
One-to-Base Coverage Algorithm for various values of cost function U , we implemented it using the open-source Player/Stage robot control system and the Boost
Graph Library (BGL). All results presented here are generated using Player 2.1.1,
Stage 2.1.1, and BGL 1.34.1. A non-convex environment (borrowed from [26]) is
specified with three robots. The free space is modeled using an occupancy grid
with 0.6m resolution, producing a lattice-like graph with all edge weights equal
to 0.6m. The 0.6m resolution is chosen so that each robot can fit in a grid cell.
120
Chapter 4. Partitioning with One-to-Base-Station Communication
One example with U = Hmin,inf is shown in Figure 4.4. In the simulation, the
robots have uniform weight assignment defined by wi =
1
3
for i ∈ {1, . . . , 3}. We
start with each robot owning the entire environment and stationed at its unique
centroid as shown in the first panel, and then proceed by choosing a random robot
to communicate with the base station at each iteration. The second panel shows
an intermediate covering of the environment before convergence to a centroidal
Voronoi partition. The third panel shows convergence of the U = Hmin portion of
the U = Hmin,inf algorithm. The fourth panel shows the Pareto-optimal partition
which is achieved after convergence of the U = Hinf portion of the U = Hmin,inf
algorithm. As can be seen, the movement of the robot relative to the third panel
is marginal, but the partition appears to be more balanced and is still centroidal
Voronoi. The cost to cover in terms of the multi-center function, H, decreases
from H = 729 to H = 728. Although the final partition and the decrease in costto-cover change only marginally in this example, the change can be much more
noticeable as is explained in the following.
Another example with U = Hmin,inf is shown in Figure 4.5. As before, the
robot’s have uniform weight assignment defined by wi =
1
3
for i ∈ {1, . . . , 3}.
The example starts with each agent owning the entire territory, the agents being
stationed at their unique centroid, and the simulation continuing with the agents
being selected at random to communicate with the base station. The second panel
121
Chapter 4. Partitioning with One-to-Base-Station Communication
Figure 4.4: Snapshots from a simulation of three robots partitioning an environment with black obstacles using the Hmin,inf One-to-base station algorithm.
The free space of the environment is modeled using the indicated occupancy grid
where each cell is a vertex in the resulting graph. The robots’ optimal coverage
position is marked by an X and the boundary of each robot’s territory drawn in
its color. Some cells are on the boundary of multiple territories: for these we draw
superimposed robot colors.
shows the convergence of the U = Hmin portion of the algorithm, which leads to
a final multi-center cost of H = 804. The third panel shows the update after the
first iteration of the U = Hinf portion of the algorithm with the green agent. The
update shows that the lower portion of the environment is getting less than optimal
coverage and is improved by moving an agent closer to that region. The fourth
panel shows the Pareto-optimal partition which is achieved after convergence of
the U = Hinf portion of the U = Hmin,inf algorithm. Notice that in this example,
the final partition is quite different from the partition achieved at the end of the
U = Hmin portion of the algorithm. The final multi-center function cost of this
partition is H = 753, which is a noticeable improvement in coverage.
Thus far we have looked at two representative examples of using the algorithm with U = Hmin,inf . These examples illustrate that, like centroidal Voronoi
122
Chapter 4. Partitioning with One-to-Base-Station Communication
Figure 4.5: Snapshots from a simulation of three robots partitioning an environment with black obstacles using the Hmin,inf One-to-base station algorithm.
Note that the initial condition is the same as in Figure 4.4, but the evolution is
different.
partitions, Pareto-optimal partitions are not necessarily unique, and that the evolution under the One-to-base station algorithm is only guaranteed to converge to
a locally optimal solution. To see how the algorithm compares in general and for
different choices of U , we simulate the algorithm with the same initial setup as
shown in both Figure 4.4 and Figure 4.5. The One-to-Base Coverage Algorithm
with U = Hmin , U = Hinf , and U = Hmin,inf is run 100 times for each choice of U .
Table 4.1 summarizes the final cost-of-coverage for each choice of U . We observe
that the One-to-Base Coverage Algorithm with U = Hmin converges to partitions
that have the same minimum cost as those attained with the algorithm using
U = Hinf or U = Hmin,inf . On the other hand, the maximum cost-to-cover with
U = Hmin can be much larger than with the other two choices of U . Of the three
algorithms, the algorithm with U = Hinf converges consistently to partitions with
the lowest coverage costs, however, as discussed earlier it is computationally the
123
Chapter 4. Partitioning with One-to-Base-Station Communication
most expensive. Finally, the algorithm with U = Hmin,inf behaves as expected,
converging on average to partitions with values similar to that of the algorithm
with U = Hinf although with a slightly larger deviation.
Algorithm
Hmin
Hinf
Hmin,inf
Min
728
728
728
Mean
746.02
730.26
730.38
Max
804
732
753
StdDev
27.74
1.92
4.91
Table 4.1: Multi-center function cost-to-cover statistics for each algorithm from
100 simulation runs.
4.3.1
Handling Dynamic Changes
Evolving overlapping coverings in the One-to-Base Coverage Algorithm enables simple handling of environmental changes along with dynamic arrivals, departures, and even the disappearance of robots. Changes in the environment along
with robot departures or disappearances can increase coverage cost, but those increases are only a transients and, with the appropriate algorithmic additions, the
system will converge in finite steps after such an event. The One-to-Base Coverage Algorithm also has the added advantage that it can account for changes in
robot performance due to changes in capability caused by potential damage to the
hardware. The following algorithmic additions address how to handle the events
described.
124
Chapter 4. Partitioning with One-to-Base-Station Communication
Environment Changes Each region in the environment is initially assigned
an importance according to the weight function φ(x). As robots explore the
environment, they may determine that certain regions are more/less important
than what was originally assigned. Robots can communicate this to the base
station at which point the base station can update φ(x).
Arrival When a new robot i communicates with the base station, it can be
assigned any initial Pi desired. Possibilities include adding all vertices within a
set distance of its initial position or assigning it just the single vertex which has
the highest coverage cost in Q.
Departure & Disappearance A robot i might announce to the base station
that it is departing, perhaps to recharge its batteries or to perform some other
task. In this situation, the base station can simply add Pi to the territory of the
next robot it talks to before executing the normal steps of the algorithm. The
disappearance or failure of a robot i can be detected if it does not communicate
with the base station for longer than ∆. If this occurs, then the departure procedure above can be triggered. Should i reappear later, it can be handled as a new
arrival or given its old territory.
125
Chapter 4. Partitioning with One-to-Base-Station Communication
Performance Malfunction of a robot i can be detected by the agent via self
diagnosis and communicated to the base station. If this malfunction causes the
robot to survey less territory, then wi will have changed, so the base station can
simply re-normalize the vector of weights w.
4.4
Summary
We have described a coverage algorithm, with corresponding cost-functions,
which uses the One-to-Base station communication architecture that drive territory ownership among a team of robots in a non-convex environment to a centroidal Voronoi partition in finite time. We have also defined the notion of Paretooptimal partition and have provided a provably correct method to reach such a
partition using the One-to-Base Coverage Algorithm. Finally, we have demonstrated the effectiveness of the algorithm through simulation, and have outlined
various ways the algorithm can be adapted to allow for dynamic changes in the
system. We have focused on dividing territory in this work, but the algorithm can
easily be combined with methods to provide a service over Q, as in [14].
In this chapter we looked at coverage control in a discrete time and space
environment with unreliable asynchronous communication. In the next chapter
126
Chapter 4. Partitioning with One-to-Base-Station Communication
we look at what happens in a continuous time and space environment, and when
there are constraints on the area owned by each agent.
127
Chapter 5
Partitioning with
Area-Constraints
This chapter is organized as follows. In section 5.1 we setup preliminary notation, introduce the concept of generalized Voronoi partitions and present our
problem in technical detail. In section 5.2 we compute some useful properties of
the objective functions of interest. In section 5.3 we state existence properties of
area-constrained generalized Voronoi partitions and present algorithms to reach
the set of area-constrained Voronoi partitions. In section 5.4 we state the main
result of our paper on centroidal equitable generalized Voronoi partitions. In section 5.6 we discuss the application of our algorithm to real systems and provide
numerical simulations. In the final section we summarize our findings.
128
Chapter 5. Partitioning with Area-Constraints
5.1
Preliminaries and problem statement
Let us have a convex compact set Q ⊂ R2 , endowed with a density function
φ : Q → R≥0 , so that the measure (or area) of a region A ⊂ Q is defined as
Z
φ(A) =
φ(q)dq,
A
provided the set A is measurable in the sense of Lebesgue. Without loss of generality, we assume that Q has unit measure, that is, φ(Q) =
R
Q
φ(q)dq = 1. Let
p1 , . . . , pn denote the positions of n robotic agents in Q. We assume that each agent
is associated with a (measurable) sub-region Wi ⊂ Q, where {Wi }ni=1 partitions Q
into sets whose interiors are pairwise disjoint. A vector can be defined to collect
the measures of the regions of a partition, as φ(W ) = [φ(W1 ), . . . , φ(Wn )]T . By our
assumptions on Q, we have
Pn
i=1
φ(Wi ) = 1. Let f : R → R be a strictly convex,
increasing, and differentiable function. Then, given n locations p = (p1 , . . . , pn )
and a partition W = (W1 , . . . , Wn ), the multicenter function is defined by
H(p, W ) =
n Z
X
i=1
f (kq − pi k)φ(q)dq.
Wi
Our goal in this work is minimizing the function H under certain constraints,
namely, that the areas of each region are fixed. Specifically, we consider constants
ci > 0 for each agent i ∈ {1, . . . , n} such that
Pn
i=1 ci
ci for every i. For brevity, we denote S = {c ∈ Rn>0 |
129
= 1 and we require φ(Wi ) =
Pn
i=1 ci
= 1}.
Chapter 5. Partitioning with Area-Constraints
Problem 7 (Multicenter optimization with area constraint). Given c ∈ S, determine the locations of the agents p = (p1 , . . . , pn ) and the partition W =
(W1 , . . . , Wn ) solving:
min H(p, W )
p,W
(5.1)
subject to φ(Wi ) = ci , i ∈ {1, . . . , n − 1}.
Note that the nth constraint φ(Wn ) = cn is omitted because it is redundant.
In order to solve this problem, we introduce a useful partitioning scheme. To
begin, we define D := {p ∈ Qn | pi 6= pj , i 6= j} as the set of disjoint positions in
Q. Then, given the function f as above, n distinct locations p ∈ D, and n scalar
weights w = (w1 , . . . , wn ), the generalized Voronoi partition of Q is the collection
of subsets V f (p, w) = (V1f (p, w), . . . , Vnf (p, w)) of Q, defined by
Vif (p, w) = {q ∈ Q | f (kq − pi k) − wi ≤ f (kq − pj k) − wj , ∀j 6= i}.
(5.2)
Generalized Voronoi partitions enjoy several important properties. First, any generalized Voronoi partition of Q is in fact a partition of Q. Second, the generalized
partition generated by (p, w) is equal to the generalized partition generated by
(p, w + α1n ), for any α ∈ R (Here, 1n is the vector in Rn whose entries are all
equal to 1). Finally, as opposed to Voronoi partitions, for generalized Voronoi
partitions we do not require that pi ∈ Vif (p, w). Thus, agents need not be located
in their partition.
From here onward, we will refer to generalized Voronoi par130
Chapter 5. Partitioning with Area-Constraints
titions simply as Voronoi partitions. Two important special cases are described
below.
Example 2 (Standard Voronoi Diagram). Given p ∈ D and n scalar weights
w = (w1 , . . . , wn ), the Standard Voronoi Diagram of Q is given by (5.2) with
w = 0. The partition is given by
ViSD (p, w) = {q ∈ Q | f (kq − pi k) ≤ f (kq − pj k)},
regardless of the choice of f (x). We call each ViSD a standard Voronoi region.
These regions are convex and have boundaries that are given by straight line segments; moreover, every generator pi is contained in its respective region ViSD .
Example 3 (Power Diagrams). Given p ∈ D and n scalar weights w = (w1 , . . . , wn ),
the power diagram of Q is given by (5.2) with f (x) = x2 . The partition is given
by
ViPD (p, w) = {q ∈ Q | kq − pi k2 − wi ≤ kq − pj k2 − wj },
and we call each Voronoi region ViPD a power cell. Note that Standard Voronoi
Diagrams are a special case of Power Diagrams, since ViPD (p, 0) = V SD (p). These
regions are convex and their boundaries that are line segments; however, it is
possible that the generators pi are not contained by their respective power cells
ViPD .
131
Chapter 5. Partitioning with Area-Constraints
B
B
A
A
C
C
D
D
E
E
Figure 5.1: The image on the left is the Standard Voronoi Partition generated
by nodes A through E. The image on the right shows the dual graph for this
partition.
We now define the dual graph of a partition {Wi }ni=1 , which will be useful
later: the node set is {1, . . . , n}, and there exists an edge {i, j} if the boundary
between agents i and j, denoted ∆i,j , has positive measure. In that case, we say
that j is a neighbor of i and we write j ∈ Ni . The dual graph of the standard
Voronoi Partition is known to be the classic Dirichlet triangulation, as illustrated
in Figure 5.1.
Before we are ready to define a second problem of interest, we first introduce
a useful definition. Given n distinct locations p = (p1 , . . . , pn ), the set of weights
w = (w1 , . . . , wn ) such that every partition has non-zero measure is defined by
U = {w ∈ Rn | φ(Vif (p, w)) > 0 ∀i}. Since V f (p, w) = V f (p, w + α1n ) for every
scalar α, this equivalence relation naturally defines equivalence classes in U : with
132
Chapter 5. Partitioning with Area-Constraints
a slight abuse of notation, in what follows we sometimes refer to such classes as
the elements of U . The following problem is a simplified version of Problem 7.
Problem 8 (Multicenter Voronoi partition optimization with area constraints).
Given c ∈ S, determine the locations of the agents p ∈ D and weights w ∈ U
solving:
min H(p, V f (p, w))
p,w
subject to φ(Vif (p, w)) = ci , i ∈ {1, . . . , n − 1}.
As a preliminary step, we should make sure that this problem has feasible
solutions: this fact is shown in Section 5.3, which also provides a method to find
a set of weights in U for every set of locations. Problems 7 and 8 are known to
be equivalent in the following sense.
Proposition 45 (Proposition V.1 in [23]). Let p ∈ Qn be the agent locations
and w ∈ U a weight assignment which satisfies the area constraint. Then, the
Voronoi partition V f (p, w) optimizes H(p, W ) among all partitions satisfying the
area constraint.
In order to derive a useful consequence of this fact, we consider the simpler
case when there is only one agent in Q: then, the multicenter function becomes
p 7→ H1 p, Q :=
Z
f (kq − pk)φ(q)dq.
Q
133
(5.3)
Chapter 5. Partitioning with Area-Constraints
Since f is strictly convex, H1 is too, and the following holds: If Q is convex, then
there is a unique minimizer of (5.3), which we denote by Ce(Q). Moreover, we have
that
∂H1 (p,Q)
∂p
= 0 at p = Ce(Q), where
∂H1 (p,Q)
∂p
=
∂
f (kq − pk)φ(q)dq
Q ∂p
R
from [13].
The Voronoi partition V f (p, w) generated by (p, w) is said to be centroidal if
Ce[Vif (p, w)] = pi ,
for all i ∈ {1, . . . , n}. This notation allows us to state the following fact [23]: for
every solution (p∗ , W ∗ ) of Problem 7, there exists a weight assignment w∗ ∈ U
such that W ∗ = V f (p∗ , w∗ ) and p∗i = Ce(Wi∗ ) for all i ∈ {1, . . . , n}. Equivalently,
the solutions to Problem 7 are centroidal Voronoi partitions whose regions have
the prescribed areas.
In the rest of this paper, we will go beyond this abstract characterization of
the optimal solutions and give an optimization algorithm which is amenable to
practical implementation.
5.2
Relevant partial derivatives
This section is devoted to compute relevant partial derivatives of the multicenter function, which shall be used to solve Problem 8 in the subsequent sections.
134
Chapter 5. Partitioning with Area-Constraints
Given p ∈ D, and w ∈ U , we define the partition of Q by V f (p, w). It is
convenient to define the Voronoi multi-center function as
(p, w) 7→ H(p, w) =
n Z
X
Vif (p,w)
i=1
f (kq − pi k)φ(q)dq,
or equivalently H(p, w) = H(p, V f (p, w)). It is also convenient to define the
generators-to-areas function as
"Z
#T
Z
(p, w) 7→M(p, w) =
φ(q)dq, . . . ,
V1f (p,w)
φ(q)dq
,
Vnf (p,w)
or equivalently M(p, w) = φ(V f (p, w)). In the rest of this section, we shall
compute the gradients of the functions H and M. In order to state our results, we
need some notation. Let ∆i,j (p, w) denote the boundary between the ith and jth
Voronoi region and ~ni,j the normal to this boundary, pointing towards region Wj .
Given locations p ∈ D and weights w ∈ U , let La (p, w) and Lbk (p, w), k ∈ {1, 2},
(k)
be the n × n matrices, whose entries ai,j and bi,j are defined by
ai,j (p, w) =


R

∂q

− ∆i,j φ(q) ∂wi · ~ni,j dq,
if i 6= j,


P

− n ai,j ,
i=1
otherwise,
(5.4)
and
(k)
bi,j (p, w) =


R


− ∆i,j φ(q)
∂q
(k)
∂pi


P

(k)
− n bi,j
,
i=1
135
· ~ni,j dq,
if i 6= j,
(5.5)
otherwise,
Chapter 5. Partitioning with Area-Constraints
where q ∈ R2 has components q (1) and q (2) . Clearly, entries ai,j and bi,j are zero if
∆i,j has zero measure.
We are now ready to state the two main results of this section, which imply
that computing the gradients of H(p, w) and the M(p, w) is spatially-distributed
over the dual graph of the Voronoi partition.
Proposition 46 (Partial derivatives of the Voronoi multi-center function). Given
(k)
p ∈ D and w ∈ U , let pi , k ∈ {1, 2}, denote the two components of pi ∈ R2 , for
i ∈ {1, . . . , n}, and define La and Lbk as in equations (5.4) and (5.5). Then, the
partial derivatives of H(p, w) are
Z
∂
∂H(p, w)
=
f (kq − p1 k)φ(q)dq, . . . ,
(k)
(k)
∂p
Vif ∂p1
Z
∂
f (kq − pn k)φ(q)dq + wT Lbk (p, w),
(k)
f
Vn ∂pn
(5.6)
∂H(p, w)
= wT La (p, w).
∂w
(5.7)
Proof. Note that we write V for V f throughout the proof and everything is done
(k)
with respect to one component of pi where pi ∈ R2 (we drop the k from pi
for
clarity). Differentiating with respect to pi we see that
∂H
=
∂pi
Z
∂
f (kq − pi k)φ(q)dq
Vi ∂pi
Z
∂q
+
f (kq − pi k)φ(q)
· ~ni,j dq
∂pi
∂Vi
XZ
∂q
+
f (kq − pj k)φ(q)
· ~nj,i dq,
∂pi
j∈N ∂Vi ∩ ∂Vj
i
136
(5.8)
Chapter 5. Partitioning with Area-Constraints
where (5.8) easily falls out from the conservation law [13, Proposition 2.23]. The
second term on the RHS is defined as
Z
f (kq − pi k)φ(q)
∂Vi
∂q
· n dq
∂pi
∂q
=
· ~ni,j dq
f (kq − pi k)φ(q)
∂p
i
∆
i,j
j∈Ni
XZ
∂q
=
(f (kq − pj k) + wi − wj )φ(q)
· ~ni,j dq,
∂p
i
∆
i,j
j∈N
XZ
i
where f (kq − pi k) = f (kq − pj k) + wi − wj holds true along the boundary ∆i,j
and is given from the definition of the Voronoi partition. Therefore, combining
the second and third terms on the RHS of (5.8) and noting that ~ni,j = −~nj,i , we
have
∂H
=
∂pi
Z
X
∂q
∂
(wi − wj )
φ(q)
f (kq − pi k)φ(q)dq +
· ~ni,j dq. (5.9)
∂p
i
∆
Vi ∂pi
i,j
j∈N
Z
i
Writing (5.9) in vector form gives
∂H(p,w)
∂p
as defined in (5.6), with matrices whose
entries are defined by (5.5). Similarly the derivative with respect to wi is given as
Z
X
∂H
∂q
=
(wi − wj )
φ(q)
· ~ni,j dq.
∂wi j∈N
∂wi
∆i,j
(5.10)
i
The vector form of this equation can easily be seen to be (5.7) with the Laplacian
matrix La defined by (5.4).
Proposition 47 (Partial derivatives of the generators-to-areas function). Given
(k)
p ∈ D and w ∈ U , let pi , k ∈ {1, 2}, denote the two components of pi ∈ R2 , for
137
Chapter 5. Partitioning with Area-Constraints
i ∈ {1, . . . , n}, and define La and Lbk as in equations (5.4) and (5.5). Then, the
partial derivatives of M(p, w) are
∂M(p, w)
= Lbk (p, w),
∂p(k)
(5.11)
∂M(p, w)
= La (p, w).
∂w
(5.12)
Proof. For clarity, we write V for V f throughout the proof. Looking at the ith
component of M(p, w) and differentiating with respect to wi , by the conservation
law [13, Proposition 2.23] we get
Z
∂
∂q
φ(q)
φ(q)dq +
· ~ni,j dq,
∂wi
Vi ∂wi
∂Vi
XZ
∂q
φ(q)
· ~ni,j dq,
=
∂wi
j∈N ∆i,j
∂Mi
=
∂wi
Z
(5.13)
i
where ∆i,j is the boundary between agents i and j. The first term on the right
hand side of (5.13) is zero since the density function is not dependent on the
weights. Similarly, the derivative of Mj with respect to wi , is given as
Z
∂q
∂
φ(q)dq +
φ(q)
· ~nj,i dq,
∂wi
Vj ∂wi
∂Vj
Z
∂q
=−
φ(q)
· ~ni,j dq,
∂wi
∆i,j
∂Mj
=
∂wi
Z
(5.14)
where we note that ~ni,j = −~nj,i . Therefore it is easily seen that the total gradient
of G with respect to the vector of weights w is given by (5.12), with La (p, w)
(k)
defined by (5.4). Similarly for pi
we get the gradient defined by (5.11) whose
matrix Lbk (p, w) is given by (5.5).
138
Chapter 5. Partitioning with Area-Constraints
The following useful Proposition follows from Proposition 47 and Proposition IV.1 in [23] and is due to the monotonicity properties of the Voronoi partition. Indeed, any increase in the weights of agent i (keeping the weights of other
agents fixed) guarantees that the area of agent i increases and the areas of the
neighboring agents decrease.
Proposition 48 (Sign-definiteness of the partial derivatives of M). Given p ∈ D
and w ∈ U , then
∂Mi (p, w)
∂Mi (p, w)
> 0 and
≤ 0, j 6= i,
∂wi
∂wj
where the second inequality is strict if and only if ∆i,j has non-zero measure.
Next, consider the dual graph of the Voronoi partition defined by (p, w) and
assign to each edge {i, j} of this graph the (i, j) entry of La (p, w), which is
strictly-negative by Proposition 48. With this definition, the matrix La (p, w)
is the Laplacian matrix naturally associated to this weighted dual graph of the
Voronoi partition defined by (p, w).
5.3
Area-constrained Voronoi partitions
In this section, we solve the problem of finding, given area constraints and
locations of the agents, suitable weights such that the Voronoi partition generated
139
Chapter 5. Partitioning with Area-Constraints
by these locations and weights satisfies the area constraint. We begin by stating
two useful results.
Proposition 49 (Existence and uniqueness of weights for area-constrained Voronoi
partitions). Define constants c ∈ S. Given p ∈ D, there exists a unique vector w∗ ∈ U , up to translation, such that {V1f (p, w∗ ), . . . , Vnf (p, w∗ )} satisfies
φ(Vif ) = ci for all i.
Proof. Existence follows from Proposition IV.4 in [23]. Before beginning the
uniqueness argument, we introduce the set Vi→j (w∗ , w) as the set of points that
move from agent i to agent j due to a change in weights from w∗ to w. For example, if agents i and j are neighbors, wj > wj∗ , and wi = wi∗ for all i 6= j, then
φ(Vi→j (w∗ , w)) > 0. That is to say that some region is transferred to agent j due
to its weight increasing. With this notation in mind, we begin the proof with a set
of weights, w∗ ∈ U , which define some arbitrary partition. Associated with this
partition are a set of non-zero areas, c∗i > 0 for i ∈ {1, . . . , n}, that correspond
to each agents region. Assume there exists another set of weights, w 6= w∗ , such
that M(p, w) = M(p, w∗ ) = c∗ . Since weights are translation invariant, we can
translate w such that w1∗ − w1 = 0. Without loss of generality (wlog), assume that
generator 2 is the neighbor of generator 1 in V f (p, w∗ ), and that w2 −w1 < w2∗ −w1∗
(w2 < w2∗ , since w 6= w∗ ) or equivalently w2 − w2∗ < w1 − w1∗ = 0. Due to the
monotonicity of the weights, w2 < w2∗ implies that part of the region that was
140
Chapter 5. Partitioning with Area-Constraints
once owned by agent 2 in V f (p, w∗ ) is either now owned by agent 1 or by some
other agent in the partition V f (p, w). Thus φ(V2→i (w∗ , w)) > 0 for some i 6= 2.
This means that agent 2 in V f (p, w) must own the space of some other agent
(or combination of agents) in order to maintain its area-constraint. This can
only happen if there exists at least one neighboring agent (wlog; agent 3) whose
weight satisfies the condition w3 − w2 < w3∗ − w2∗ . Thus it must be the case that
w3 − w3∗ < w2 − w2∗ < w1 − w1∗ and w3 < w3∗ . Since every agent has at least
one neighbor, we can continue in this fashion of ordering agents, until we reach
agent k such that wj − wj∗ < wk − wk∗ , where j > k, cannot be satisfied, and so
φ(Vkf (p, w)) < c∗k . Thus, the set of weights that satisfy the area-constraint for a
generalized Voronoi partition are unique, up to translation.
Given this result, for any given set of area constraints c ∈ S, we may formally
define the map wac : D → U as p 7→ wac (p), such that
Pn
i=1 (wac (p))i
= 0 and
φ(Vif (p, wac (p))) = ci for all i.
Proposition 50 (Smoothness of mapping from positions to weights). The map
p 7→ wac (p) is continuously differentiable.
Proof. The proof makes use of the implicit function theorem in conjunction with a
modified mapping of M(p, w), to be described later, to show continuous differentiability of wac (p). Let c ∈ S denote the vector of areas for the constraint surface.
141
Chapter 5. Partitioning with Area-Constraints
Then given the mapping wac (p), we have that M(p, wac (p)) = c. Since Voronoi
partitions are translation invariant with respect to weights, w, we can define any
Voronoi partition as a function of n − 1 weights, keeping the nth weight constant
at zero. Define w˜ ≡ [w1 , . . . , wn−1 ]T , and define the modified mapping (p, w)
˜ 7→
hR
iT
R
n−1
˜
˜
M(p, w)
˜ ∈R
of M(p, w) by M(p, w)
˜ = V f (p,w)
φ(q)dq, . . . , V f (p,w)
φ(q)dq .
˜
˜
1
n−1
˜
Define the mapping p 7→ w˜ac (p) such that M(p,
w˜ac (p)) = c˜, where we set c˜i = ci
for i ∈ {1, . . . , n − 1}. Note that this is sufficient to define the constraint surface
since the nth constraint in M(p, w) and the nth weight wn are both redundant.
For clarity, calculations in the rest of the proof are done with respect to one com˜
ponent of pi , where pi ∈ R2 . Differentiating M(p,
w)
˜ = c˜ with respect to w˜ and p
we obtain
˜
∂ M(p,
w)
˜
˜ a (p, w),
=L
˜
∂ w˜
˜
∂ M(p,
w)
˜
˜ b (p, w),
=L
˜
∂p
˜ b (p, w)
˜ a (p, w)
where L
˜ is defined as Lb (p, w) with the nth row removed and L
˜ is
defined as the Laplacian matrix La (p, w) with the nth column and row removed
(Note Lb (p, w) is dependent on which component of pi we choose). Proposition 48
guarantees that the Laplacian of the Voronoi dual graph is always well-defined,
˜ a (p, w˜ac (p)) ∈ Rn−1×n−1 and is full rank [40, Corollary 6.2.27]. Theretherefore L
˜
fore, since M(p,
w)
˜ is continuously differentiable and
142
˜ w)
∂ M(p,
˜
∂w
˜
is invertible, then by
Chapter 5. Partitioning with Area-Constraints
the implicit function theorem we have that w˜ac (p) is continuously differentiable.
Finally, note that the mapping wac (p) can be written as [w˜ac (p)T , wac,n (p)]T , where
wac,n (p) is a constant value of zero. Since w
˜ac (p) and wac,n (p) are continuously
differentiable, so is wac (p).
We now present an algorithm to compute wac (p) for a specific area-constraint.
Given p ∈ D and w ∈ U , the area-constraint cost function for the Voronoi partition
generated by (p, w) is defined as
Z
J (p, w) = n log
φ(q)dq −
Q
= n log (φ(Q)) −
n
X
i=1
n
X
Z
ci log
φ(q)dq
Vif (p,w)
f
ci log φ Vi (p, w)
(5.15)
i=1
where ci for i ∈ {1, . . . , n} are strictly positive constants. Note that n log (φ(Q)) =
0 when φ(Q) = 1. The following result extends [71, Theorem 3.7] to (generalized)
Voronoi partitions, and has the added property of being better conditioned numerically.
Theorem 51 (Gradient of the area-constraint cost function). Let ∆i,j denote the
boundary between the ith and j th Voronoi region and ~ni,j the normal vector along
that boundary. Define constants c ∈ S. Given p ∈ D and w ∈ U , we have
X
∂
J p, w =
Ξi,j cj φ(Vif ) − ci φ(Vjf ) ,
∂wi
j∈N
i
143
(5.16)
Chapter 5. Partitioning with Area-Constraints
where
Ξi,j =
Z
1
φ(Vif )φ(Vjf )
∆i,j
∂q
· ~ni,j φ(q)dq,
∂wi
(5.17)
so that
(i) every w generating an area-constrained Voronoi partition is a critical point
of the function w 7→ J p, w), and
(ii) every solution to the negative gradient flow
w˙ i = −
∂
J p, w ,
∂wi
(5.18)
converges asymptotically to wac (p), yielding an area-constrained Voronoi partition such that φ(Vif ) = ci .
Proof. Let w 7→ J p, w) be a candidate Lyapunov function. First, we check
that J is continuously differentiable. Using (5.13) and (5.14) from the proof of
Proposition 47 and the chain rule, we quickly have (5.16) with coefficients defined
by (5.17). Therefore given p ∈ D, we have that J is continuously differentiable
with respect to the weights w ∈ U . Second, we see with w˙ i defined according
P
to (5.18) that J˙ = ni=1
let yi =
1
,
φ(Vif (p,w))
∂J
w˙
∂wi i
≤ 0. To determine when J˙ is identically zero,
and rewrite (5.16) and (5.17) in vector notation to get the
following:
∂
J p, w = y T diag ([c1 , . . . , cn ]) La (p, w),
∂w
144
Chapter 5. Partitioning with Area-Constraints
where La (p, w) is the Laplacian matrix defined by (5.4). Let x = y T diag ([c1 , . . . , cn ]),
then
∂
J
∂w
p, w is identically zero when x = α1Tn for any constant α. Given the
constraints of the system this can only happen when yi = c1i , or equivalently when
φ Vif (p, w) = ci for all i ∈ {1, . . . , n}. Thus the invariant set for (5.18) is such
P
f
that for φ Vi (p, w) = ci all i and ni=1 ci = 1. By these observations, we have
proved claim (i). Third, we have to show that trajectories are bounded. From
the gradient descent law (5.18) we deduce that the measures of each agent are
bounded away from zero. Indeed, if the measure of an agent’s region were to
approach zero, then the function J would grow unbounded; this is impossible
because we know J is monotonically non-increasing. Notice that the measures
of the agents depend on the weights, and it is not hard to verify that the sum
of weights stays constant. Hence, if a weight were to become very large, another
weight would become arbitrarily small, which would cause a region to vanish. This
contradicts the fact that the measures are bounded: therefore, the weights must
also be bounded. After these three observations, we can invoke LaSalle Invariance
Principle and deduce that the weights converge to the set of weights w such that
φ(Vif (p, w)) = ci for all i. Additionally, since the sum of the weights is constant
and the vector of weights that satisfy the area constraint is unique by Proposition 49, we conclude that the weights converge to the vector of weights w∗ such
that
Pn
i=1
wi∗ =
Pn
i=1
wi (0) and φ(Vif (p, w∗ )) = ci for all i, proving claim (ii).
145
Chapter 5. Partitioning with Area-Constraints
We now specialize the gradient (5.18) to the case of Example 3.
Example 4 (Gradient flow for equitable power diagrams). Define the partition of
the region Q according to (5.2) with f (x) = x2 and let ci =
From [71] we have that
∂q
∂wi
· ~ni,j =
1
kpi −pj k
1
n
for all i in (5.15).
so then the gradient flow (5.18) is given
by
w˙ i = −
X
j∈Ni
1
1
−
PD
φ(Vj ) φ(ViPD )
Z
∆i,j
φ(q)
dq.
kpi − pj k
The vector of weights w converges to the value such that φ ViPD (p, w) = φ VjPD (p, w)
for all i 6= j.
5.4
Centroidal area-constrained Voronoi partitions
Given constants c ∈ S, and given p ∈ D, it is convenient to define the areaconstrained Voronoi partition generated by p as
Vacf (p) = V f (p, wac (p)),
f
f
with Voronoi regions Vac,i
(p), i ∈ {1, . . . , n} such that φ Vac,i
(p) = ci for all i.
The associated area-constrained multicenter function is given by
p 7→ H
p, Vacf (p)
=
n Z
X
i=1
f
Vac,i
(p)
146
f (kq − pi k)φ(q)dq,
Chapter 5. Partitioning with Area-Constraints
or equivalently by p 7→ H p, wac (p) . We are now ready for the main result of this
section.
Theorem 52 (Gradient of the area-constrained multicenter function). Given p ∈
D,
∂
H p, Vacf (p) =
∂pi
Z
f
Vac,i
(p)
∂
f (kq − pi k)φ(q)dq,
∂pi
(5.19)
so that
(i) every p generating a centroidal area-constrained Voronoi partition is a crit
ical point of the function p 7→ H p, Vacf (p) , and
(ii) every solution to the negative gradient flow
Z
p˙i = −
f
Vac,i
(p)
∂
f (kq − pi k)φ(q)dq
∂pi
(5.20)
converges asymptotically to the set of centroidal area-constrained Voronoi
partitions.
Proof. Let H(p, wac (p)) = H(p, V f (p, wac (p))), the Voronoi multicenter function
restricted to the area-constraint surface, be our candidate Lyapunov function. By
differentiating with respect to p we obtain
∂H(p, wac ) ∂H(p, wac ) dwac
∂
H(p, wac ) =
+
,
∂p
∂p
∂w
dp
#
f
f
∂H1 (pn , Vac,n
(p))
(p))
∂H1 (p1 , Vac,1
=
,...,
∂p1
∂pn
dwac
T
+ w Lb (p, w) + wT La (p, w)
,
dp
147
(5.21)
Chapter 5. Partitioning with Area-Constraints
where La (p, w) and Lb (p, w) are defined by (5.4) and (5.5), respectively. Differentiating M(p, wac (p)) = c with respect to p we obtain
∂M(p, wac (p)) ∂M(p, wac (p)) dwac (p)
+
= 0,
∂p
∂w
dp
ac (p)
ac (p)
= 0, therefore La (p, w) dwdp
= −Lb (p, w). Subthat is, Lb (p, w) + La (p, w) dwdp
∂
∂
H(p, V f (p, wac )) = ∂p
stituting this equality into (5.21) it follows that ∂p
H(p, wac ) =
f
f
∂H1 (p1 ,Vac,1
(p))
∂H1 (pn ,Vac,n
(p))
,
.
.
.
,
. Therefore H(p, V f (p, wac )) is continuously dif∂p1
∂pn
ferentiable with respect to p, and its critical points are characterized, proving claim
(i). For each agent the trajectories under (5.20) point towards the centers of their
region, and since Q is convex and compact this gives that the trajectories stay in Q
and are bounded. We must also show that the agents maintain distinct locations.
If two agents i and j have the same weight (wj = wi ), then relative to each other
they generate a Standard Voronoi partition (Example 2) and the center for each
agent stays in the same region as the agent, therefore the agents can not collide.
Now we look at the case when the weights are different. Without loss of generality
let wj > wi and assume that agent j approaches agent i. From (5.2) we have that
for q ∈ Q, the region of agent j satisfies f (kq − pj k) − wj + wi ≤ f (kq − pi k).
If pj and pi are close enough, all points in a neighborhood of pi belong to region
j. This implies that the measure of region i is zero; this is impossible since the
flow stays along the area-constraint surface. Therefore, agents can not collide and
agent locations remain distinct along the flow. Under control law (5.20) we have
148
Chapter 5. Partitioning with Area-Constraints
that
˙ V f (p, wac )) =
H(p,
f
n
X
∂H1 (pi , Vac,i
(p))
∂pi
i=1
=−
n
X
p˙i
f
∂H1 (pi , Vac,i
(p))
∂pi
i=1
!2
.
By LaSalle’s Invariance Principle the positions p converge to the invariant set of
f
positions such that pi = Ce[Vac,i
(p)] for all i. Therefore, the positions converge to
the set of centroidal area-constrained Voronoi partitions and claim (ii) is proved.
There are some interesting points worth noting. First, assuming that the set of
centroidal area-constrained Voronoi partitions is finite, the positions and weights
converge to one of the partitions in that set. Second, the gradient descent (5.20)
is not guaranteed to find the global minimum. Finally, the gradient restricted to
the constraint surface is formally the same as the reduced gradient as defined in
nonlinear programming [57].
We now specialize the gradient (5.19) to the case of Example 3.
Example 5 (Constrained gradient flow for power diagrams). Let the partition of
the region Q be defined according to (5.2) with f (x) = x2 . Given any powercell
A of Q the center is given by Ce[A] =
1
φ(A)
R
A
qφ(q)dq. Thus for power diagrams,
Ce is equivalent to the well known expression for center of mass of a region.
149
Chapter 5. Partitioning with Area-Constraints
From [13] we have that
PD (p)
∂H1 p,Veq,i
∂pi
=
1
n
PD
] − pi , therefore the (scaled,
Ce[Vac,i
negative) gradient flow (5.20) is given by
PD
p˙i = Ce[Vac,i
(p)] − pi .
PD
We easily see that H(p, VacPD ) is minimized when pi = Ce[Vac,i
(p)] for all i ∈
{1, . . . , n}.
5.5
Simultaneous change of agent positions and
weights
The previous gradient descent laws (5.18) and (5.20), are designed to find the
area-constraint surface and reach the center of a area-constrained region, respectively. In this section, we introduce a distributed algorithm which achieves both
tasks simultaneously. As before, for the desired constraint surface we choose constants c ∈ S. Then, given p ∈ D and w ∈ U , the simultaneous gradient algorithm
is given by
Z
p˙i = −
w˙ i = −
Vif (p,w)
X
∂
f (kq − pi k)φ(q)dq,
∂pi
Ξi,j cj φ(Vif ) − ci φ(Vjf ) ,
j∈Ni
150
(5.22)
Chapter 5. Partitioning with Area-Constraints
where
Ξi,j =
1
φ(Vif )φ(Vjf )
Z
∆i,j
∂q
· ~ni,j φ(q)dq.
∂wi
The proposed control law is a natural combination of laws (5.18) and (5.20). If all
the weights are initialized to the same value w, and if w˙ i is set to zero, then (5.22)
reduces to the continuous-time Lloyd algorithm presented in [13]. Simulations
show that the proposed law does in fact converge to the set of centroidal areaconstrained Voronoi partitions; however, a proof is not currently available.
We now specialize the gradient (5.22) to the case of Example 3.
Example 6 (Gradient flow for equitable power diagrams). Define the partition of
the region Q according to (5.2) with f (x) = x2 and let ci = n1 for all i in (5.15).
∂H1 p,ViPD (p)
1
PD
=
From [13] we have that
Ce[V
]
−
p
, therefore the (scaled,
i
i
∂pi
n
negative) gradient flow (5.22) is given by
p˙i = Ce[ViPD (p)] − pi ,
Z
X 1
1
φ(q)
w˙ i = −
−
dq.
PD
PD
φ(Vj ) φ(Vi )
∆i,j kpi − pj k
j∈N
i
PD
We easily see that the stationary set for this system is achieved when pi = Ce[Vac,i
(p)]
for all i ∈ {1, . . . , n} and the vector of weights w converges to the value such that
φ ViPD (p, w) = φ VjPD (p, w) for all i 6= j.
151
Chapter 5. Partitioning with Area-Constraints
5.6
Implementation and simulations
In this section, we discuss the practical implementation of the control law
(5.20) and compare it against the control law (5.22) using representative simulation examples. Writing the area-constrained gradient flow (5.20), we assume
that we always stay on the constraint surface and thus move along this surface
continuously. In order to put this law into practice, we would need an explicit
formula to instantaneously compute the weights of the current Voronoi partition,
as functions of the generator locations. Since such a formula is not available, we
instead have to rely on system (5.18) in order to determine the weights. We then
design an implementation which alternates dynamics (5.20) and (5.18). Assuming
that the agents start at a feasible configuration, they move their locations according to (5.20) for a small time duration δ, while keeping the weights fixed. After
this amount of time, the area constraint is not satisfied: we then let the weights
evolve according to (5.18), while the locations are fixed, until we are within the
proximity of the area-constraint surface.
Some points in this implementation require attention. First, if the positions are
allowed to move too much, the regions can become undefined (i.e., have measure
zero), therefore care must be taken to make sure this does not happen by selecting
a sufficiently small δ. Second, care must also be taken to insure that the agent
152
Chapter 5. Partitioning with Area-Constraints
location do not collide (i.e., agent locations must remain distinct). If they do,
the step size δ should be reduced in order to avoid collision. Third, in order to
drive the system exactly back to the constraint surface, we would need to bring
the dynamics of the weights to convergence, which would take an infinitely long
time: in simulations, convergence is approximated up to truncation error. In spite
of these difficulties, in our simulations we have found that the algorithm is not
sensitive to how far the agents deviate from the area-constraint surface (provided
measures stay non-zero) during each movement step: in all our experiments, the
algorithm converges to a centroidal area-constrained Voronoi diagram.
An illustrative example of the performance of the algorithm is presented in
Figure 5.2. In the simulation, 10 agents have been randomly placed in a square
region Q, where the density function φ(x) is constant. The region Q is to be
partitioned according to (5.2) with f (x) = x2 , that is, as a power diagram. We
define area-constraint surface such that φ(Vif ) =
Pni
i=1
i
φ(Q) for all i ∈ {1, . . . , 10};
which means that if i < j, then φ(Vif ) < φ(Vjf ). The gradient (5.20) is followed
during steps of duration δ = 0.1s. The first panel shows the initial condition of
the system at T = 0, where each agent has been randomly placed and the corresponding weights which generate the area-constrained partition determined. In
the second panel, at T = δ, the agents have moved in the gradient direction which
causes each region to have a different area (the agents moved off the constraint
153
Chapter 5. Partitioning with Area-Constraints
6
2
6
2
6
2
9
9
9
8
8
7
4
7
4
8
4
5
5
7
3
1
3
5
3
1
1
Figure 5.2: Simulation of 9 agents partitioning a square environment with uniform density using the iterative gradient algorithm.
surface of equitable area). The last panel shows the final state of the system at
T = 54.2s. Each agent is at its region’s centroid and the regions have the desired
measure.
Next we observe the performance of control law (5.22), which simultaneously
optimizes the weights and the positions. We set the initial weights of the system
to zero, but we keep the same initial positions and area-constraint requirements
described above. The first panel of Figure 5.3 shows that we do in fact start
with the same initial positions as the previous control algorithm, and since the
weights are zero, the partition is the Standard Voronoi Diagram; the second panel
shows how the position trajectories evolve over time, and the final panel shows
the system’s final configuration. Figure 5.4 shows better how the areas of each
agent’s region evolves over time; the second panel shows how each agent’s position
converges to its regions centroid. It is interesting to observe that given the same
154
Chapter 5. Partitioning with Area-Constraints
2
6
6
2
9
9
8
7
4
8
4
5
7
5
3
3
1
1
Figure 5.3: Simulation of 9 agents partitioning a square environment with uniform density using the simultaneous gradient algorithm.
initial conditions, the iterative algorithm and control law (5.22) converge to the
same configuration. In fact, we did not come across a case where this did not
happen.
5.7
Summary
In this chapter we studied the problem of how to optimally deploy a set of
agents over a convex workspace while each agent maintains a pre-specified area.
We have designed a provably correct, spatially-distributed continuous-time policy
that solves this optimization problem. We proposed a method for implementation
of the control policy and demonstrated its effectiveness in simulation.
155
Chapter 5. Partitioning with Area-Constraints
0.25
0.2
||pi − ci ||
Area of Region
0.2
0.15
0.1
0.05
0
0.15
0.1
0.05
0
10
20
30
Time (s)
40
50
0
0
10
20
30
Time (s)
40
50
Figure 5.4: Area (left) and position (right) trajectories for 9 agents partitioning a square environment with uniform density using the simultaneous gradient
algorithm.
156
Chapter 6
Conclusion and Future Work
Coordination in multi-agent networks is becoming more relevant as tasks become increasingly complex and need to be distributed. There are a myriad of
applications for such networks, each with their own complex and interesting problems. In this thesis we have tackled two important problems in particular.
First, we looked at the problem of surveillance. In particular, we looked
at stochastic surveillance, which has applications when the strategy has unpredictability or predefined visit frequency requirements; in many surveillance problems there are important regions that needed to be visited more frequently than
others. We showed how to formulate such problems and how to optimally find
stochastic surveillance strategies.
Second, we studied the problem of deployment and territory partitioning with
a mobile robotic network. This problem has many variations which depend on
the type of communication law used and constraints applied. We looked at two
157
Chapter 6. Conclusion and Future Work
cases. We began by looking at the case where there is sporadic, asynchronous
communication between agents in the network and a central base-station. This
sees applications when continuous line of sight communication is not readily possible; underwater vehicles and spacecraft are two examples of this. Then, we looked
at the case when there is continuous communication with neighboring agents and
there is a constraint on the amount of coverage each agent must provide. This
arises in applications where workload needs to be distributed in a predefined way,
such as distributing workload amongst mobile cell towers.
6.1
Summary
We started in Chapter 2 by formally formulating the stochastic surveillance
problem over a graph. We modeled our surveillance strategy as a Markov chain
described by a transition matrix and showed that minimizing the hitting time
of that Markov chain gives the optimal strategy for a certain class of problems.
Using the hitting time of the Markov chain, we showed for reversible transition
matrices that the hitting time can be formulated as an SDP, and thus a globally
optimal solution, if it exists, can always be found. Then, we extended the notion
of hitting time to graphs with travel distances and showed, for reversible transition
matrices, that the weighted hitting time can also be formulated as a SDP. Since
158
Chapter 6. Conclusion and Future Work
the weighted hitting time accounts for travel distances, it proved to be far superior
than other reversible Markov chain based surveillance strategies.
In Chapter 3 we further developed the notion of hitting time, and were able to
extend hitting time to multiple random walkers, coining it the group hitting time.
As a consequence, our results also allowed us to determine the hitting time between
any set of nodes, a quantity that could previously only be calculated for (1) a single
random walker with (2) a very specialized structure. These results also allowed us
to extend the weighted hitting time by finding analytic expressions for the pairwise
hitting times for a graph with travel distances, however, this was only possible for
the single random walker case. Finally, we provided extensive simulations showing
the random walks that generate optimal group hitting times can have a predictable
structure for some graph topologies, and highly unpredictable structure for others.
In Chapter 4 we started looking into problems in coverage control. We considered a discrete environment and modeled the environment by a graph as was done
in the previous two chapters. In this chapter we defined the one-to-base station
communication protocol and provided an algorithm that, using the one-to-base
protocol, converged to centroidal Voronoi partitions. We also introduced the notion of Pareto-optimal partitions and provided a provable method to converge to
those partitions. In addition, we discussed various real-world application cases,
and how to handle each using the algorithms presented.
159
Chapter 6. Conclusion and Future Work
Finally, in Chapter 5 we studied the area-constrained coverage problem over
a convex workspace. Using generalized Voronoi partitions, we showed that when
each agent can talk to its territory neighbor, that we can design a provably correct,
spatially-distributed continuous-time policy that converges to the set of centroidal
area-constrained partitions.
6.2
Future Work
This work leaves open various directions for further research.
Hitting time For the single agent weighted hitting time in Chapter 2, it was
only optimized over the transition matrix. It would be interesting to see the implications of optimizing both the weight matrix and transition matrix simultaneously.
This can have the interpretation of optimizing the “capacity” or “resistance” of
the graph, a topic in optimization which is of independent interest [34].
For the multiple Markov chains scenario, a clear extension is to consider the
case of heterogeneous travel times similar to what was done in Chapter 2. In addition, although we provide a method for determining the hitting time of multiple
random walkers, in general, it can be difficult to compute when the size of the
graph and number of agents increases. It would be of practical interest to find
a formulation which is computationally less expensive or a method in which the
160
Chapter 6. Conclusion and Future Work
group hitting time can be calculated in a distributed way. On a related note, we
found in Chapter 3 that the minimal group hitting time is linked to the underlying
graph topology of each random walker. It would be of interest to find a way to
incorporate spatial partitioning algorithms, such as those in Chapter 4 or [26], to
find graph partitions/coverings that generate lower hitting times. Finally, given
the maximum pairwise hitting time of a Markov chain, there exists bounds on the
cover time for multiple copies of that Markov chain running in parallel [27, 3]. It
would be interesting to see if our results can be leveraged to extend those bounds
to multiple heterogenous Markov chains running in parallel.
Coverage Control For the one-to-base algorithm it would be worthwhile to
adapt the algorithm to allow for area-constrained partitions similar to the work
done in Chapter 5. Also, we would like to extend the One-to-base Coverage
Algorithm to other communication settings (e.g., directional or pair-wise gossip)
to take advantage of the notion of Pareto-optimal partitions.
For area-constrained partitions (Chapter 5) our main approach is based on
alternating phases, during which we either improve the objective function, or enforce the area constraint. In contrast to our main solution, we would like to
prove convergence of the proposed policy in which agents converge to the constraint surface while simultaneously optimizing the coverage problem. So far, the
161
Chapter 6. Conclusion and Future Work
effectiveness of this policy has been observed in simulations. Second, our policy requires synchronous and reliable communication along the edges of the dual
graph associated to the Voronoi partition. It would be of practical interest to relax
this requirement, using asynchronous, event-based, or unreliable communication
as in Chapter 4 or [12, 62]. In addition, our approach is based on the assumption
that the environment we partition is convex, and finding policies that work over
non-convex environments would be of great practical use: works in this direction
include [16] and [26]. Finally, we assume that the density function φ is known to
the agents, which may be hard to satisfy in practice; several papers have recently
appeared to overcome this assumption, including [63] and [74].
162
Bibliography
[1] A. Albert. Conditions for positive and nonnegative definiteness in terms of
pseudoinverses. SIAM Journal on Applied Mathematics, 17(2):434–440, 1969.
[2] D. Aldous and J. A. Fill. Reversible Markov Chains and Random Walks on
Graphs, 2002. Unfinished monograph, recompiled 2014, available at http:
//www.stat.berkeley.edu/~aldous/RWG/book.html.
[3] N. Alon, C. Avin, M. Koucky, G. Kozma, Z. Lotker, and M. R. Tuttle. Many
random walks are faster than one. Combinatorics, Probability and Computing,
20, 7 2011.
¨
[4] D. A. Anisi, P. Ogren,
and X. Hu. Cooperative minimum time surveillance
with multiple ground vehicles. IEEE Transactions on Automatic Control,
55(12):2679–2691, 2010.
[5] E. Bakshy, I. Rosenn, C. Marlow, and L. Adamic. The role of social networks
in information diffusion. In Int. Conference on World Wide Web, pages 519–
528, Lyon, France, 2012.
[6] S. Bhattacharya, R. Ghrist, and V. Kumar. Multi-robot coverage and exploration in non-Euclidean metric spaces. In Algorithmic Foundations of
Robotics X, volume 86, pages 245–262. Springer, 2013.
[7] T. H. Blackwell and J. S. Kaufman. Response time effectiveness: Comparison
of response time and survival in an urban emergency medical services system.
Academic Emergency Medicine, 9(4):288–295, 2002.
[8] B. Boˇsansk´
y, V. Lis´
y, M. Jakob, and M. Pˇechouˇcek. Computing timedependent policies for patrolling games with mobile targets. In International
Conference on Autonomous Agents and Multiagent Systems, pages 989–996,
2011.
163
Bibliography
[9] S. Boyd, P. Diaconis, and L. Xiao. Fastest mixing Markov chain on a graph.
SIAM Review, 46(4):667–689, 2004.
[10] S. Boyd and L. Vandenberghe. Convex Optimization. Cambridge University
Press, 2004.
[11] A. Z. Broder and A. R. Karlin. Bounds on the cover time. Journal of
Theoretical Probability, 2(1):101–120, 1989.
[12] F. Bullo, R. Carli, and P. Frasca. Gossip coverage control for robotic networks: Dynamical systems on the space of partitions. SIAM Journal on
Control and Optimization, 50(1):419–447, 2012.
[13] F. Bullo, J. Cort´es, and S. Mart´ınez. Distributed Control of Robotic Networks.
Princeton University Press, 2009.
[14] F. Bullo, E. Frazzoli, M. Pavone, K. Savla, and S. L. Smith. Dynamic vehicle
routing for robotic systems. Proceedings of the IEEE, 99(9):1482–1504, 2011.
[15] P. Burgisser, M. Clausen, and A. Shokrollahi. Algebraic Complexity Theory.
Springer, 1997.
[16] J. G. Carlsson, E. Carlsson, and R. Devulapalli. Balancing workloads for service vehicles over a geographic territory. IEEE/RSJ Int. Conf. on Intelligent
Robots & Systems, Nov. 2013. To appear.
[17] J. G. Carlsson and R. Devulapalli. Shadow prices in territory division. University of Minnesota. Available online at http://menet.umn.edu/~jgc/, 2013.
[18] M. Catral, S. J. Kirkland, M. Neumann, and N.-S. Sze. The Kemeny constant for finite homogeneous ergodic Markov chains. Journal of Scientific
Computing, 45(1-3):151–166, 2010.
[19] L. Chen and J. Leneutre. A game theoretical framework on intrusion detection in heterogeneous networks. IEEE Transactions on Information Forensics
and Security, 4(2):165–178, June 2009.
[20] J. Clark and R. Fierro. Mobile robotic sensors for perimeter detection and
tracking. ISA Transactions, 46(1):3–13, 2007.
[21] C. Cooper, A. Frieze, and T. Radzik. Multiple random walks and interacting
particle systems. In S. Albers, A. Marchetti-Spaccamela, Y. Matias, S. Nikoletseas, and W. Thomas, editors, Automata, Languages and Programming,
164
Bibliography
volume 5556 of Lecture Notes in Computer Science, pages 399–410. Springer,
2009.
[22] C. Cooper, A. M. Frieze, and T. Radzik. Multiple random walks in random
regular graphs. SIAM Journal on Discrete Mathematics, 23(4):1738–1761,
2009.
[23] J. Cort´es. Coverage optimization and spatial load balancing by robotic sensor
networks. IEEE Transactions on Automatic Control, 55(3):749–754, 2010.
[24] J. Cort´es, S. Mart´ınez, T. Karatas, and F. Bullo. Coverage control for
mobile sensing networks. IEEE Transactions on Robotics and Automation,
20(2):243–255, 2004.
[25] P. G. Doyle and J. L. Snell. Random Walks and Electric Networks. Mathematical Association of America, 1984.
[26] J. W. Durham, R. Carli, P. Frasca, and F. Bullo. Discrete partitioning
and coverage control for gossiping robots. IEEE Transactions on Robotics,
28(2):364–378, 2012.
[27] K. Efremenko and O. Reingold. How well do random walks parallelize? In
I. Dinur, K. Jansen, J. Naor, and J. Rolim, editors, Approximation, Randomization, and Combinatorial Optimization. Algorithms and Techniques, volume
5687 of Lecture Notes in Computer Science, pages 476–489. Springer, 2009.
[28] W. Ellens, F. M. Spieksma, P. V. Mieghem, A. Jamakovic, and R. E. Kooij.
Effective graph resistance. Linear Algebra and its Applications, 435(10):2491–
2506, 2011.
[29] Y. Elmaliach, N. Agmon, and G. A. Kaminka. Multi-robot area patrol under frequency constraints. Annals of Mathematics and Artificial Intelligence,
57(3-4):293–320, 2009.
[30] Y. Elmaliach, A. Shiloni, and G. A. Kaminka. A realistic model of frequencybased multi-robot polyline patrolling. In International Conference on Autonomous Agents, pages 63–70, Estoril, Portugal, May 2008.
[31] R. Els¨asser and T. Sauerwald. Tight bounds for the cover time of multiple
random walks. Theoretical Computer Science, 412(24):2623–2641, 2011.
[32] U. Feige. A tight upper bound on the cover time for random walks on graphs.
Random Structures & Algorithms, 6(1):51–54, 1995.
165
Bibliography
[33] E. Fiorelli, N. E. Leonard, P. Bhatta, D. A. Paley, R. Bachmayer, and D. M.
Fratantoni. Multi-AUV control and adaptive sampling in Monterey Bay.
IEEE Journal of Oceanic Engineering, 31(4):935–948, 2006.
[34] A. Ghosh, S. Boyd, and A. Saberi. Minimizing effective resistance of a graph.
SIAM Review, 50(1):37–66, 2008.
[35] B. Golden, S. Raghavan, and E. Wasil. The Vehicle Routing Problem: Latest
Advances and New Challenges, volume 43 of Operations Research/Computer
Science Interfaces. Springer, 2008.
[36] J. Grace and J. Baillieul. Stochastic strategies for autonomous robotic surveillance. In IEEE Conf. on Decision and Control and European Control Conference, pages 2200–2205, Seville, Spain, Dec. 2005.
[37] M. Grant and S. Boyd. CVX: Matlab software for disciplined convex programming, version 2.1. http://cvxr.com/cvx, Oct. 2014.
[38] Z. Han, A. L. Swindlehurst, and K. J. R. Liu. Smart deployment/movement of
unmanned air vehicle to improve connectivity in MANET. In IEEE Wireless
Communications and Networking Conference, pages 252 –257, Apr. 2006.
[39] W. K. Hastings. Monte Carlo sampling methods using Markov chains and
their applications. Biometrika, 57(1):97–109, 1970.
[40] R. A. Horn and C. R. Johnson. Matrix Analysis. Cambridge University Press,
1985.
[41] R. A. Horn and C. R. Johnson. Topics in Matrix Analysis. Cambridge
University Press, 1994.
[42] J. J. Hunter. Generalized inverses and their application to applied probability
problems. Linear Algebra and its Applications, 45:157–198, 1982.
[43] J. J. Hunter. Mathematical Techniques of Applied Probability, volume 1 of
Discrete Time Models: Basic Theory. Academic Press, 1983.
[44] J. J. Hunter. Mathematical Techniques of Applied Probability, volume 2 of
Discrete Time Models: Techniques and Applications. Academic Press, 1983.
[45] J. J. Hunter. The role of Kemeny’s constant in properties of Markov chains.
Communications in Statistics - Theory and Methods, 43(7):1309–1321, 2014.
166
Bibliography
[46] A. K. Jain, M. N. Murty, and P. J. Flynn. Data clustering: A review. ACM
Computing Surveys, 31(3):264–323, 1999.
[47] J. G. Kemeny and J. L. Snell. Finite Markov Chains. Springer, 1976.
[48] D. B. Kingston, R. W. Beard, and R. S. Holt. Decentralized perimeter surveillance using a team of UAVs. IEEE Transactions on Robotics, 24(6):1394–
1404, 2008.
[49] S. Kirkland. Fastest expected time to mixing for a Markov chain on a directed
graph. Linear Algebra and its Applications, 433(11-12):1988–1996, 2010.
[50] D. J. Klein and M. Randi´c. Resistance distance. Journal of Mathematical
Chemistry, 12(1):81–95, 1993.
[51] J. Leskovec, D. Chakrabarti, J. Kleinberg, and C. Faloutsos. Realistic, mathematically tractable graph generation and evolution, using Kronecker multiplication. In Knowledge Discovery in Databases (PKDD), pages 133–145.
Springer, 2005.
[52] J. Leskovec and C. Faloutsos. Scalable modeling of real graphs using Kronecker multiplication. In Int. Conference on Machine Learning, pages 497–
504, Corvalis, OR, USA, 2007.
[53] M. Levene and G. Loizou. Kemeny’s constant and the random surfer. The
American Mathematical Monthly, 109(8):741–745, 2002.
[54] D. A. Levin, Y. Peres, and E. L. Wilmer. Markov Chains and Mixing Times.
American Mathematical Society, 2009.
[55] Y. Y. Li and L. E. Parker. Intruder detection using a wireless sensor network
with an intelligent mobile robot response. In IEEE SoutheastCon, pages
37–42, Huntsville, AL, USA, Apr. 2008.
[56] S. P. Lloyd. Least squares quantization in PCM. IEEE Transactions on
Information Theory, 28(2):129–137, 1982. Presented at the 1957 Institute for
Mathematical Statistics Meeting.
[57] D. G. Luenberger. Linear and Nonlinear Programming. Addison-Wesley, 2
edition, 1984.
[58] A. Macwan, G. Nejat, and B. Benhabib. Optimal deployment of robotic
teams for autonomous wilderness search and rescue. In IEEE/RSJ Int. Conf.
167
Bibliography
on Intelligent Robots & Systems, pages 4544–4549, San Francisco, CA, USA,
Sept. 2011.
[59] M. Mahdian and Y. Xu. Stochastic Kronecker graphs. Random Structures &
Algorithms, 38(4):453–466, 2011.
[60] C. D. Meyer. Matrix Analysis and Applied Linear Algebra. SIAM, 2001.
[61] Y. Nonaka, H. Ono, K. Sadakane, and M. Yamashita. The hitting and cover
times of Metropolis walks. Theoretical Computer Science, 411(16):1889–1894,
2010.
[62] C. Nowzari and J. Cort´es. Self-triggered coordination of robotic networks for
optimal deployment. Automatica, 48(6):1077–1087, 2012.
[63] J. L. Ny and G. J. Pappas. Adaptive deployment of mobile robotic networks.
IEEE Transactions on Automatic Control, 58(3):654–666, 2013.
[64] A. Okabe, B. Boots, K. Sugihara, and S. N. Chiu. Spatial Tessellations:
Concepts and Applications of Voronoi Diagrams. Wiley Series in Probability
and Statistics. John Wiley & Sons, 2 edition, 2000.
[65] J. L. Palacios. Closed-form formulas for Kirchhoff index. International Journal of Quantum Chemistry, 81(2):135–140, 2001.
[66] J. L. Palacios. On hitting times of random walks on trees. Statistics &
Probability Letters, 79(2):234–236, 2009.
[67] J. L. Palacios and P. Tetali. A note on expected hitting times for birth and
death chains. Statistics & Probability Letters, 30(2):119–125, 1996.
[68] F. Pasqualetti, A. Franchi, and F. Bullo. On cooperative patrolling: Optimal trajectories, complexity analysis and approximation algorithms. IEEE
Transactions on Robotics, 28(3):592–606, 2012.
[69] A. Patcha and J. Park. A game theoretic approach to modeling intrusion
detection in mobile ad hoc networks. In Information Assurance Workshop.
Proceedings from the Fifth Annual IEEE SMC, pages 280–284, June 2004.
[70] R. Patel, P. Agharkar, and F. Bullo. Robotic surveillance and Markov chains
with minimal first passage time. IEEE Transactions on Automatic Control,
May 2014. To appear.
168
Bibliography
[71] M. Pavone, A. Arsie, E. Frazzoli, and F. Bullo. Distributed algorithms for
environment partitioning in mobile robotic networks. IEEE Transactions on
Automatic Control, 56(8):1834–1848, 2011.
[72] A. Pereira, H. Heidarsson, C. Oberg, D. Caron, B. Jones, and G. Sukhatme.
A communication framework for cost-effective operation of AUVs in coastal
regions. In A. Howard, K. Iagnemma, and A. Kelly, editors, Field and Service
Robotics, volume 62 of Tracts in Advanced Robotics, pages 433–442. Springer,
2010.
[73] L. C. A. Pimenta, V. Kumar, R. C. Mesquita, and G. A. S. Pereira. Sensing and coverage for a network of heterogeneous robots. In IEEE Conf. on
Decision and Control, pages 3947–3952, Canc´
un, M´exico, Dec. 2008.
[74] M. Schwager, D. Rus, and J. J. Slotine. Decentralized, adaptive coverage
control for networked robots. International Journal of Robotics Research,
28(3):357–375, 2009.
[75] S. D. Servetto and G. Barrenechea. Constrained random walks on random
graphs: routing algorithms for large scale wireless sensor networks. In Proceedings of the 1st ACM International Workshop on Wireless Sensor Networks and Applications, pages 12–21. ACM, 2002.
[76] C. Seshadhri, A. Pinar, and T. G. Kolda. An in-depth analysis of stochastic
Kronecker graphs. Journal of the Association for Computing Machinery,
60(2):13:1–13:32, 2013.
[77] R. C. Shah, S. Roy, S. Jain, and W. Brunette. Data MULEs: modeling
and analysis of a three-tier architecture for sparse sensor networks. Ad Hoc
Networks, 1(2-3):215–233, 2003.
[78] R. N. Smith, Y. Chao, P. P. Li, D. A. Caron, B. H. Jones, and G. S. Sukhatme.
Planning and implementing trajectories for autonomous underwater vehicles
to track evolving ocean processes based on predictions from a regional ocean
model. International Journal of Robotics Research, 29(12):1475–1497, 2010.
[79] K. Srivastava, D. M. Stipanovi`c, and M. W. Spong. On a stochastic robotic
surveillance problem. In IEEE Conf. on Decision and Control, pages 8567–
8574, Shanghai, China, Dec. 2009.
[80] V. Srivastava, F. Pasqualetti, and F. Bullo. Stochastic surveillance strategies
for spatial quickest detection. International Journal of Robotics Research,
32(12):1438–1458, 2013.
169
Bibliography
[81] S. Susca, S. Mart´ınez, and F. Bullo. Monitoring environmental boundaries
with a robotic sensor network. IEEE Transactions on Control Systems Technology, 16(2):288–296, 2008.
[82] P. Tetali. Random walks and the effective resistance of networks. Journal of
Theoretical Probability, 4(1):101–109, 1991.
[83] V. V. Vazirani. Approximation Algorithms. Springer, 2001.
[84] P. M. Weichsel. The Kronecker product of graphs. Proceedings of the American Mathematical Society, 13(1):47–52, 1962.
[85] X. Wu and Z. Liu. How community structure influences epidemic spread
in social networks. Physica A: Statistical Mechanics and its Applications,
387(2):623–630, 2008.
[86] P. R. Wurman, R. D’Andrea, and M. Mountz. Coordinating hundreds of
cooperative, autonomous vehicles in warehouses. AI Magazine, 29(1):9–20,
2008.
170